nsz Git - libfirm/blob - ir/lower/lower_dw.c

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief   Lower double word operations, i.e. 64bit -> 32bit, 32bit -> 16bit etc.
   9  * @date    8.10.2004
  10  * @author  Michael Beck
  11  */
  12 #include "config.h"
  13
  14 #include <string.h>
  15 #include <stdlib.h>
  16 #include <stdbool.h>
  17 #include <assert.h>
  18
  19 #include "be.h"
  20 #include "error.h"
  21 #include "lowering.h"
  22 #include "irnode_t.h"
  23 #include "irnodeset.h"
  24 #include "irgraph_t.h"
  25 #include "irmode_t.h"
  26 #include "iropt_t.h"
  27 #include "irgmod.h"
  28 #include "tv_t.h"
  29 #include "dbginfo_t.h"
  30 #include "iropt_dbg.h"
  31 #include "irflag_t.h"
  32 #include "firmstat.h"
  33 #include "irgwalk.h"
  34 #include "ircons.h"
  35 #include "irflag.h"
  36 #include "iroptimize.h"
  37 #include "irtools.h"
  38 #include "debug.h"
  39 #include "set.h"
  40 #include "pmap.h"
  41 #include "pdeq.h"
  42 #include "irdump.h"
  43 #include "array_t.h"
  44 #include "irpass_t.h"
  45 #include "lower_dw.h"
  46
  47 /** A map from (op, imode, omode) to Intrinsic functions entities. */
  48 static set *intrinsic_fkt;
  49
  50 /** A map from (imode, omode) to conv function types. */
  51 static set *conv_types;
  52
  53 /** A map from a method type to its lowered type. */
  54 static pmap *lowered_type;
  55
  56 /** A map from a builtin type to its lower and higher type. */
  57 static pmap *lowered_builtin_type_high;
  58 static pmap *lowered_builtin_type_low;
  59
  60 /** The types for the binop and unop intrinsics. */
  61 static ir_type *binop_tp_u, *binop_tp_s, *unop_tp_u, *unop_tp_s, *tp_s, *tp_u;
  62
  63 static ir_nodeset_t created_mux_nodes;
  64
  65 /** the debug handle */
  66 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  67
  68 /**
  69  * An entry in the (op, imode, omode) -> entity map.
  70  */
  71 typedef struct op_mode_entry {
  72         const ir_op   *op;    /**< the op */
  73         const ir_mode *imode; /**< the input mode */
  74         const ir_mode *omode; /**< the output mode */
  75         ir_entity     *ent;   /**< the associated entity of this (op, imode, omode) triple */
  76 } op_mode_entry_t;
  77
  78 /**
  79  * An entry in the (imode, omode) -> tp map.
  80  */
  81 typedef struct conv_tp_entry {
  82         const ir_mode *imode; /**< the input mode */
  83         const ir_mode *omode; /**< the output mode */
  84         ir_type       *mtd;   /**< the associated method type of this (imode, omode) pair */
  85 } conv_tp_entry_t;
  86
  87 enum lower_flags {
  88         MUST_BE_LOWERED = 1,  /**< graph must be lowered */
  89         CF_CHANGED      = 2,  /**< control flow was changed */
  90 };
  91
  92 /**
  93  * The lower environment.
  94  */
  95 typedef struct lower_dw_env_t {
  96         lower64_entry_t **entries;     /**< entries per node */
  97         ir_graph      *irg;
  98         struct obstack obst;           /**< an obstack holding the temporary data */
  99         ir_tarval *tv_mode_bytes;      /**< a tarval containing the number of bytes in the lowered modes */
 100         pdeq      *waitq;              /**< a wait queue of all nodes that must be handled later */
 101         ir_node  **lowered_phis;       /**< list of lowered phis */
 102         ir_mode   *high_signed;        /**< doubleword signed type */
 103         ir_mode   *high_unsigned;      /**< doubleword unsigned type */
 104         ir_mode   *low_signed;         /**< word signed type */
 105         ir_mode   *low_unsigned;       /**< word unsigned type */
 106         ident     *first_id;           /**< .l for little and .h for big endian */
 107         ident     *next_id;            /**< .h for little and .l for big endian */
 108         const lwrdw_param_t *params;   /**< transformation parameter */
 109         unsigned flags;                /**< some flags */
 110         unsigned n_entries;            /**< number of entries */
 111 } lower_dw_env_t;
 112
 113 static lower_dw_env_t *env;
 114
 115 static void lower_node(ir_node *node);
 116
 117 /**
 118  * Create a method type for a Conv emulation from imode to omode.
 119  */
 120 static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
 121 {
 122         conv_tp_entry_t key, *entry;
 123         ir_type *mtd;
 124
 125         key.imode = imode;
 126         key.omode = omode;
 127         key.mtd   = NULL;
 128
 129         entry = set_insert(conv_tp_entry_t, conv_types, &key, sizeof(key), hash_ptr(imode) ^ hash_ptr(omode));
 130         if (! entry->mtd) {
 131                 int n_param = 1, n_res = 1;
 132
 133                 if (imode == env->high_signed || imode == env->high_unsigned)
 134                         n_param = 2;
 135                 if (omode == env->high_signed || omode == env->high_unsigned)
 136                         n_res = 2;
 137
 138                 /* create a new one */
 139                 mtd = new_type_method(n_param, n_res);
 140
 141                 /* set param types and result types */
 142                 n_param = 0;
 143                 if (imode == env->high_signed) {
 144                         if (env->params->little_endian) {
 145                                 set_method_param_type(mtd, n_param++, tp_u);
 146                                 set_method_param_type(mtd, n_param++, tp_s);
 147                         } else {
 148                                 set_method_param_type(mtd, n_param++, tp_s);
 149                                 set_method_param_type(mtd, n_param++, tp_u);
 150                         }
 151                 } else if (imode == env->high_unsigned) {
 152                         set_method_param_type(mtd, n_param++, tp_u);
 153                         set_method_param_type(mtd, n_param++, tp_u);
 154                 } else {
 155                         ir_type *tp = get_type_for_mode(imode);
 156                         set_method_param_type(mtd, n_param++, tp);
 157                 }
 158
 159                 n_res = 0;
 160                 if (omode == env->high_signed) {
 161                         if (env->params->little_endian) {
 162                                 set_method_res_type(mtd, n_res++, tp_u);
 163                                 set_method_res_type(mtd, n_res++, tp_s);
 164                         } else {
 165                                 set_method_res_type(mtd, n_res++, tp_s);
 166                                 set_method_res_type(mtd, n_res++, tp_u);
 167                         }
 168                 } else if (omode == env->high_unsigned) {
 169                         set_method_res_type(mtd, n_res++, tp_u);
 170                         set_method_res_type(mtd, n_res++, tp_u);
 171                 } else {
 172                         ir_type *tp = get_type_for_mode(omode);
 173                         set_method_res_type(mtd, n_res++, tp);
 174                 }
 175                 entry->mtd = mtd;
 176         } else {
 177                 mtd = entry->mtd;
 178         }
 179         return mtd;
 180 }
 181
 182 /**
 183  * Add an additional control flow input to a block.
 184  * Patch all Phi nodes. The new Phi inputs are copied from
 185  * old input number nr.
 186  */
 187 static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
 188 {
 189         int i, arity = get_Block_n_cfgpreds(block);
 190         ir_node **in;
 191
 192         assert(nr < arity);
 193
 194         NEW_ARR_A(ir_node *, in, arity + 1);
 195         for (i = 0; i < arity; ++i)
 196                 in[i] = get_Block_cfgpred(block, i);
 197         in[i] = cf;
 198
 199         set_irn_in(block, i + 1, in);
 200
 201         foreach_out_edge(block, edge) {
 202                 ir_node *phi = get_edge_src_irn(edge);
 203                 if (!is_Phi(phi))
 204                         continue;
 205
 206                 for (i = 0; i < arity; ++i)
 207                         in[i] = get_irn_n(phi, i);
 208                 in[i] = in[nr];
 209                 set_irn_in(phi, i + 1, in);
 210         }
 211 }
 212
 213 /**
 214  * Add an additional control flow input to a block.
 215  * Patch all Phi nodes. The new Phi inputs are copied from
 216  * old input from cf tmpl.
 217  */
 218 static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
 219 {
 220         int i, arity = get_Block_n_cfgpreds(block);
 221         int nr = 0;
 222
 223         for (i = 0; i < arity; ++i) {
 224                 if (get_Block_cfgpred(block, i) == tmpl) {
 225                         nr = i;
 226                         break;
 227                 }
 228         }
 229         assert(i < arity);
 230         add_block_cf_input_nr(block, nr, cf);
 231 }
 232
 233 /**
 234  * Return the "operational" mode of a Firm node.
 235  */
 236 static ir_mode *get_irn_op_mode(ir_node *node)
 237 {
 238         switch (get_irn_opcode(node)) {
 239         case iro_Load:
 240                 return get_Load_mode(node);
 241         case iro_Store:
 242                 return get_irn_mode(get_Store_value(node));
 243         case iro_Div:
 244                 return get_irn_mode(get_Div_left(node));
 245         case iro_Mod:
 246                 return get_irn_mode(get_Mod_left(node));
 247         case iro_Cmp:
 248                 return get_irn_mode(get_Cmp_left(node));
 249         default:
 250                 return get_irn_mode(node);
 251         }
 252 }
 253
 254 /**
 255  * Walker, prepare the node links and determine which nodes need to be lowered
 256  * at all.
 257  */
 258 static void prepare_links(ir_node *node)
 259 {
 260         ir_mode         *mode = get_irn_op_mode(node);
 261         lower64_entry_t *link;
 262
 263         if (mode == env->high_signed || mode == env->high_unsigned) {
 264                 unsigned idx = get_irn_idx(node);
 265                 /* ok, found a node that will be lowered */
 266                 link = OALLOCZ(&env->obst, lower64_entry_t);
 267
 268                 if (idx >= env->n_entries) {
 269                         /* enlarge: this happens only for Rotl nodes which is RARELY */
 270                         unsigned old   = env->n_entries;
 271                         unsigned n_idx = idx + (idx >> 3);
 272
 273                         ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
 274                         memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
 275                         env->n_entries = n_idx;
 276                 }
 277                 env->entries[idx] = link;
 278                 env->flags |= MUST_BE_LOWERED;
 279         } else if (is_Conv(node)) {
 280                 /* Conv nodes have two modes */
 281                 ir_node *pred = get_Conv_op(node);
 282                 mode = get_irn_mode(pred);
 283
 284                 if (mode == env->high_signed || mode == env->high_unsigned) {
 285                         /* must lower this node either but don't need a link */
 286                         env->flags |= MUST_BE_LOWERED;
 287                 }
 288                 return;
 289         } else if (is_Call(node)) {
 290                 /* Special case:  If the result of the Call is never used, we won't
 291                  * find a Proj with a mode that potentially triggers MUST_BE_LOWERED
 292                  * to be set.  Thus, if we see a call, we check its result types and
 293                  * decide whether MUST_BE_LOWERED has to be set.
 294                  */
 295                 ir_type *tp = get_Call_type(node);
 296                 size_t   n_res, i;
 297
 298                 n_res = get_method_n_ress(tp);
 299                 for (i = 0; i < n_res; ++i) {
 300                         ir_type *rtp = get_method_res_type(tp, i);
 301
 302                         if (is_Primitive_type(rtp)) {
 303                                 ir_mode *rmode = get_type_mode(rtp);
 304
 305                                 if (rmode == env->high_signed || rmode == env->high_unsigned) {
 306                                         env->flags |= MUST_BE_LOWERED;
 307                                 }
 308                         }
 309                 }
 310         }
 311 }
 312
 313 lower64_entry_t *get_node_entry(ir_node *node)
 314 {
 315         unsigned idx = get_irn_idx(node);
 316         assert(idx < env->n_entries);
 317         return env->entries[idx];
 318 }
 319
 320 void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
 321 {
 322         lower64_entry_t *entry = get_node_entry(old);
 323         entry->low_word  = new_low;
 324         entry->high_word = new_high;
 325 }
 326
 327 ir_mode *ir_get_low_unsigned_mode(void)
 328 {
 329         return env->low_unsigned;
 330 }
 331
 332 /**
 333  * Translate a Constant: create two.
 334  */
 335 static void lower_Const(ir_node *node, ir_mode *mode)
 336 {
 337         ir_graph  *irg      = get_irn_irg(node);
 338         dbg_info  *dbg      = get_irn_dbg_info(node);
 339         ir_mode   *low_mode = env->low_unsigned;
 340         ir_tarval *tv       = get_Const_tarval(node);
 341         ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
 342         ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
 343         ir_tarval *tv_shrs  = tarval_shrs_unsigned(tv, get_mode_size_bits(low_mode));
 344         ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
 345         ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);
 346
 347         ir_set_dw_lowered(node, res_low, res_high);
 348 }
 349
 350 /**
 351  * Translate a Load: create two.
 352  */
 353 static void lower_Load(ir_node *node, ir_mode *mode)
 354 {
 355         ir_mode    *low_mode = env->low_unsigned;
 356         ir_graph   *irg = get_irn_irg(node);
 357         ir_node    *adr = get_Load_ptr(node);
 358         ir_node    *mem = get_Load_mem(node);
 359         ir_node    *low;
 360         ir_node    *high;
 361         ir_node    *proj_m;
 362         dbg_info   *dbg;
 363         ir_node    *block = get_nodes_block(node);
 364         ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
 365                                  ? cons_volatile : cons_none;
 366
 367         if (env->params->little_endian) {
 368                 low  = adr;
 369                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 370         } else {
 371                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 372                 high = adr;
 373         }
 374
 375         /* create two loads */
 376         dbg    = get_irn_dbg_info(node);
 377         low    = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
 378         proj_m = new_r_Proj(low, mode_M, pn_Load_M);
 379         high   = new_rd_Load(dbg, block, proj_m, high, mode, volatility);
 380
 381         foreach_out_edge_safe(node, edge) {
 382                 ir_node *proj = get_edge_src_irn(edge);
 383                 if (!is_Proj(proj))
 384                         continue;
 385
 386                 switch (get_Proj_proj(proj)) {
 387                 case pn_Load_M:         /* Memory result. */
 388                         /* put it to the second one */
 389                         set_Proj_pred(proj, high);
 390                         break;
 391                 case pn_Load_X_except:  /* Execution result if exception occurred. */
 392                         /* put it to the first one */
 393                         set_Proj_pred(proj, low);
 394                         break;
 395                 case pn_Load_res: {       /* Result of load operation. */
 396                         ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
 397                         ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
 398                         ir_set_dw_lowered(proj, res_low, res_high);
 399                         break;
 400                 }
 401                 default:
 402                         assert(0 && "unexpected Proj number");
 403                 }
 404                 /* mark this proj: we have handled it already, otherwise we might fall
 405                  * into out new nodes. */
 406                 mark_irn_visited(proj);
 407         }
 408 }
 409
 410 /**
 411  * Translate a Store: create two.
 412  */
 413 static void lower_Store(ir_node *node, ir_mode *mode)
 414 {
 415         ir_graph              *irg;
 416         ir_node               *block, *adr, *mem;
 417         ir_node               *low, *high, *proj_m;
 418         dbg_info              *dbg;
 419         ir_node               *value = get_Store_value(node);
 420         const lower64_entry_t *entry = get_node_entry(value);
 421         ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
 422                                    ? cons_volatile : cons_none;
 423         (void) mode;
 424
 425         assert(entry);
 426
 427         if (! entry->low_word) {
 428                 /* not ready yet, wait */
 429                 pdeq_putr(env->waitq, node);
 430                 return;
 431         }
 432
 433         irg = get_irn_irg(node);
 434         adr = get_Store_ptr(node);
 435         mem = get_Store_mem(node);
 436         block = get_nodes_block(node);
 437
 438         if (env->params->little_endian) {
 439                 low  = adr;
 440                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 441         } else {
 442                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 443                 high = adr;
 444         }
 445
 446         /* create two Stores */
 447         dbg    = get_irn_dbg_info(node);
 448         low    = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
 449         proj_m = new_r_Proj(low, mode_M, pn_Store_M);
 450         high   = new_rd_Store(dbg, block, proj_m, high, entry->high_word, volatility);
 451
 452         foreach_out_edge_safe(node, edge) {
 453                 ir_node *proj = get_edge_src_irn(edge);
 454                 if (!is_Proj(proj))
 455                         continue;
 456
 457                 switch (get_Proj_proj(proj)) {
 458                 case pn_Store_M:         /* Memory result. */
 459                         /* put it to the second one */
 460                         set_Proj_pred(proj, high);
 461                         break;
 462                 case pn_Store_X_except:  /* Execution result if exception occurred. */
 463                         /* put it to the first one */
 464                         set_Proj_pred(proj, low);
 465                         break;
 466                 default:
 467                         assert(0 && "unexpected Proj number");
 468                 }
 469                 /* mark this proj: we have handled it already, otherwise we might fall into
 470                  * out new nodes. */
 471                 mark_irn_visited(proj);
 472         }
 473 }
 474
 475 /**
 476  * Return a node containing the address of the intrinsic emulation function.
 477  *
 478  * @param method  the method type of the emulation function
 479  * @param op      the emulated ir_op
 480  * @param imode   the input mode of the emulated opcode
 481  * @param omode   the output mode of the emulated opcode
 482  * @param env     the lower environment
 483  */
 484 static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
 485                                       ir_mode *imode, ir_mode *omode)
 486 {
 487         symconst_symbol sym;
 488         ir_entity *ent;
 489         op_mode_entry_t key, *entry;
 490
 491         key.op    = op;
 492         key.imode = imode;
 493         key.omode = omode;
 494         key.ent   = NULL;
 495
 496         entry = set_insert(op_mode_entry_t, intrinsic_fkt, &key, sizeof(key),
 497                                 hash_ptr(op) ^ hash_ptr(imode) ^ (hash_ptr(omode) << 8));
 498         if (! entry->ent) {
 499                 /* create a new one */
 500                 ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
 501
 502                 assert(ent && "Intrinsic creator must return an entity");
 503                 entry->ent = ent;
 504         } else {
 505                 ent = entry->ent;
 506         }
 507         sym.entity_p = ent;
 508         return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
 509 }
 510
 511 /**
 512  * Translate a Div.
 513  *
 514  * Create an intrinsic Call.
 515  */
 516 static void lower_Div(ir_node *node, ir_mode *mode)
 517 {
 518         ir_node  *left   = get_Div_left(node);
 519         ir_node  *right  = get_Div_right(node);
 520         ir_node  *block  = get_nodes_block(node);
 521         dbg_info *dbgi   = get_irn_dbg_info(node);
 522         ir_type  *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 523         ir_mode  *opmode = get_irn_op_mode(node);
 524         ir_node  *addr   = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 525         ir_node  *in[4];
 526         ir_node  *call;
 527         ir_node  *resproj;
 528
 529         if (env->params->little_endian) {
 530                 in[0] = get_lowered_low(left);
 531                 in[1] = get_lowered_high(left);
 532                 in[2] = get_lowered_low(right);
 533                 in[3] = get_lowered_high(right);
 534         } else {
 535                 in[0] = get_lowered_high(left);
 536                 in[1] = get_lowered_low(left);
 537                 in[2] = get_lowered_high(right);
 538                 in[3] = get_lowered_low(right);
 539         }
 540         call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
 541         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 542         set_irn_pinned(call, get_irn_pinned(node));
 543
 544         foreach_out_edge_safe(node, edge) {
 545                 ir_node *proj = get_edge_src_irn(edge);
 546                 if (!is_Proj(proj))
 547                         continue;
 548
 549                 switch (get_Proj_proj(proj)) {
 550                 case pn_Div_M:         /* Memory result. */
 551                         /* reroute to the call */
 552                         set_Proj_pred(proj, call);
 553                         set_Proj_proj(proj, pn_Call_M);
 554                         break;
 555                 case pn_Div_X_regular:
 556                         set_Proj_pred(proj, call);
 557                         set_Proj_proj(proj, pn_Call_X_regular);
 558                         break;
 559                 case pn_Div_X_except:
 560                         set_Proj_pred(proj, call);
 561                         set_Proj_proj(proj, pn_Call_X_except);
 562                         break;
 563                 case pn_Div_res:
 564                         if (env->params->little_endian) {
 565                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 566                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 567                                 ir_set_dw_lowered(proj, res_low, res_high);
 568                         } else {
 569                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 570                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 571                                 ir_set_dw_lowered(proj, res_low, res_high);
 572                         }
 573                         break;
 574                 default:
 575                         assert(0 && "unexpected Proj number");
 576                 }
 577                 /* mark this proj: we have handled it already, otherwise we might fall into
 578                  * out new nodes. */
 579                 mark_irn_visited(proj);
 580         }
 581 }
 582
 583 /**
 584  * Translate a Mod.
 585  *
 586  * Create an intrinsic Call.
 587  */
 588 static void lower_Mod(ir_node *node, ir_mode *mode)
 589 {
 590         ir_node  *left   = get_Mod_left(node);
 591         ir_node  *right  = get_Mod_right(node);
 592         dbg_info *dbgi   = get_irn_dbg_info(node);
 593         ir_node  *block  = get_nodes_block(node);
 594         ir_type  *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 595         ir_mode  *opmode = get_irn_op_mode(node);
 596         ir_node  *addr   = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 597         ir_node  *in[4];
 598         ir_node  *call;
 599         ir_node  *resproj;
 600
 601         if (env->params->little_endian) {
 602                 in[0] = get_lowered_low(left);
 603                 in[1] = get_lowered_high(left);
 604                 in[2] = get_lowered_low(right);
 605                 in[3] = get_lowered_high(right);
 606         } else {
 607                 in[0] = get_lowered_high(left);
 608                 in[1] = get_lowered_low(left);
 609                 in[2] = get_lowered_high(right);
 610                 in[3] = get_lowered_low(right);
 611         }
 612         call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
 613         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 614         set_irn_pinned(call, get_irn_pinned(node));
 615
 616         foreach_out_edge_safe(node, edge) {
 617                 ir_node *proj = get_edge_src_irn(edge);
 618                 if (!is_Proj(proj))
 619                         continue;
 620
 621                 switch (get_Proj_proj(proj)) {
 622                 case pn_Mod_M:         /* Memory result. */
 623                         /* reroute to the call */
 624                         set_Proj_pred(proj, call);
 625                         set_Proj_proj(proj, pn_Call_M);
 626                         break;
 627                 case pn_Div_X_regular:
 628                         set_Proj_pred(proj, call);
 629                         set_Proj_proj(proj, pn_Call_X_regular);
 630                         break;
 631                 case pn_Mod_X_except:
 632                         set_Proj_pred(proj, call);
 633                         set_Proj_proj(proj, pn_Call_X_except);
 634                         break;
 635                 case pn_Mod_res:
 636                         if (env->params->little_endian) {
 637                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 638                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 639                                 ir_set_dw_lowered(proj, res_low, res_high);
 640                         } else {
 641                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 642                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 643                                 ir_set_dw_lowered(proj, res_low, res_high);
 644                         }
 645                         break;
 646                 default:
 647                         assert(0 && "unexpected Proj number");
 648                 }
 649                 /* mark this proj: we have handled it already, otherwise we might fall
 650                  * into out new nodes. */
 651                 mark_irn_visited(proj);
 652         }
 653 }
 654
 655 /**
 656  * Translate a binop.
 657  *
 658  * Create an intrinsic Call.
 659  */
 660 static void lower_binop(ir_node *node, ir_mode *mode)
 661 {
 662         ir_node  *left  = get_binop_left(node);
 663         ir_node  *right = get_binop_right(node);
 664         dbg_info *dbgi  = get_irn_dbg_info(node);
 665         ir_node  *block = get_nodes_block(node);
 666         ir_graph *irg   = get_irn_irg(block);
 667         ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 668         ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
 669         ir_node  *in[4];
 670         ir_node  *call;
 671         ir_node  *resproj;
 672
 673         if (env->params->little_endian) {
 674                 in[0] = get_lowered_low(left);
 675                 in[1] = get_lowered_high(left);
 676                 in[2] = get_lowered_low(right);
 677                 in[3] = get_lowered_high(right);
 678         } else {
 679                 in[0] = get_lowered_high(left);
 680                 in[1] = get_lowered_low(left);
 681                 in[2] = get_lowered_high(right);
 682                 in[3] = get_lowered_low(right);
 683         }
 684         call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
 685         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 686         set_irn_pinned(call, get_irn_pinned(node));
 687
 688         if (env->params->little_endian) {
 689                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 690                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 691                 ir_set_dw_lowered(node, res_low, res_high);
 692         } else {
 693                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 694                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 695                 ir_set_dw_lowered(node, res_low, res_high);
 696         }
 697 }
 698
 699 static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
 700 {
 701         if (get_irn_mode(node) == dest_mode)
 702                 return node;
 703         return new_r_Conv(block, node, dest_mode);
 704 }
 705
 706 /**
 707  * Moves node and all predecessors of node from from_bl to to_bl.
 708  * Does not move predecessors of Phi nodes (or block nodes).
 709  */
 710 static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
 711 {
 712         int i, arity;
 713
 714         /* move this node */
 715         set_nodes_block(node, to_bl);
 716
 717         /* move its Projs */
 718         if (get_irn_mode(node) == mode_T) {
 719                 foreach_out_edge(node, edge) {
 720                         ir_node *proj = get_edge_src_irn(edge);
 721                         if (!is_Proj(proj))
 722                                 continue;
 723                         move(proj, from_bl, to_bl);
 724                 }
 725         }
 726
 727         /* We must not move predecessors of Phi nodes, even if they are in
 728          * from_bl. (because these are values from an earlier loop iteration
 729          * which are not predecessors of node here)
 730          */
 731         if (is_Phi(node))
 732                 return;
 733
 734         /* recursion ... */
 735         arity = get_irn_arity(node);
 736         for (i = 0; i < arity; i++) {
 737                 ir_node *pred      = get_irn_n(node, i);
 738                 ir_mode *pred_mode = get_irn_mode(pred);
 739                 if (get_nodes_block(pred) == from_bl)
 740                         move(pred, from_bl, to_bl);
 741                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
 742                         ir_node *pred_low  = get_lowered_low(pred);
 743                         ir_node *pred_high = get_lowered_high(pred);
 744                         if (get_nodes_block(pred_low) == from_bl)
 745                                 move(pred_low, from_bl, to_bl);
 746                         if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
 747                                 move(pred_high, from_bl, to_bl);
 748                 }
 749         }
 750 }
 751
 752 /**
 753  * We need a custom version of part_block_edges because during transformation
 754  * not all data-dependencies are explicit yet if a lowered nodes users are not
 755  * lowered yet.
 756  * We can fix this by modifying move to look for such implicit dependencies.
 757  * Additionally we have to keep the proj_2_block map updated
 758  */
 759 static ir_node *part_block_dw(ir_node *node)
 760 {
 761         ir_graph *irg        = get_irn_irg(node);
 762         ir_node  *old_block  = get_nodes_block(node);
 763         int       n_cfgpreds = get_Block_n_cfgpreds(old_block);
 764         ir_node **cfgpreds   = get_Block_cfgpred_arr(old_block);
 765         ir_node  *new_block  = new_r_Block(irg, n_cfgpreds, cfgpreds);
 766
 767         /* old_block has no predecessors anymore for now */
 768         set_irn_in(old_block, 0, NULL);
 769
 770         /* move node and its predecessors to new_block */
 771         move(node, old_block, new_block);
 772
 773         /* move Phi nodes to new_block */
 774         foreach_out_edge_safe(old_block, edge) {
 775                 ir_node *phi = get_edge_src_irn(edge);
 776                 if (!is_Phi(phi))
 777                         continue;
 778                 set_nodes_block(phi, new_block);
 779         }
 780         return old_block;
 781 }
 782
 783 typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
 784                                     ir_node *left, ir_node *right,
 785                                     ir_mode *mode);
 786
 787 static void lower_shr_helper(ir_node *node, ir_mode *mode,
 788                              new_rd_shr_func new_rd_shrs)
 789 {
 790         ir_node  *right         = get_binop_right(node);
 791         ir_node  *left          = get_binop_left(node);
 792         ir_mode  *shr_mode      = get_irn_mode(node);
 793         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 794         ir_mode  *low_unsigned  = env->low_unsigned;
 795         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 796         ir_graph *irg           = get_irn_irg(node);
 797         ir_node  *left_low      = get_lowered_low(left);
 798         ir_node  *left_high     = get_lowered_high(left);
 799         dbg_info *dbgi          = get_irn_dbg_info(node);
 800         ir_node  *lower_block;
 801         ir_node  *block;
 802         ir_node  *cnst;
 803         ir_node  *andn;
 804         ir_node  *cmp;
 805         ir_node  *cond;
 806         ir_node  *proj_true;
 807         ir_node  *proj_false;
 808         ir_node  *phi_low;
 809         ir_node  *phi_high;
 810         ir_node  *lower_in[2];
 811         ir_node  *phi_low_in[2];
 812         ir_node  *phi_high_in[2];
 813
 814         /* this version is optimized for modulo shift architectures
 815          * (and can't handle anything else) */
 816         if (modulo_shift != get_mode_size_bits(shr_mode)
 817                         || modulo_shift2<<1 != modulo_shift) {
 818                 panic("Shr lowering only implemented for modulo shift shr operations");
 819         }
 820         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 821                 panic("Shr lowering only implemented for power-of-2 modes");
 822         }
 823         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 824         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 825                 panic("Shr lowering only implemented for two-complement modes");
 826         }
 827
 828         block = get_nodes_block(node);
 829
 830         /* if the right operand is a 64bit value, we're only interested in the
 831          * lower word */
 832         if (get_irn_mode(right) == env->high_unsigned) {
 833                 right = get_lowered_low(right);
 834         } else {
 835                 /* shift should never have signed mode on the right */
 836                 assert(get_irn_mode(right) != env->high_signed);
 837                 right = create_conv(block, right, low_unsigned);
 838         }
 839
 840         lower_block = part_block_dw(node);
 841         env->flags |= CF_CHANGED;
 842         block = get_nodes_block(node);
 843
 844         /* add a Cmp to test if highest bit is set <=> whether we shift more
 845          * than half the word width */
 846         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 847         andn       = new_r_And(block, right, cnst, low_unsigned);
 848         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 849         cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
 850         cond       = new_rd_Cond(dbgi, block, cmp);
 851         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 852         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 853
 854         /* the true block => shift_width < 1word */
 855         {
 856                 /* In theory the low value (for 64bit shifts) is:
 857                  *    Or(High << (32-x)), Low >> x)
 858                  * In practice High << 32-x will fail when x is zero (since we have
 859                  * modulo shift and 32 will be 0). So instead we use:
 860                  *    Or(High<<1<<~x, Low >> x)
 861                  */
 862                 ir_node *in[1]        = { proj_true };
 863                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 864                 ir_node *res_high     = new_rd_shrs(dbgi, block_true, left_high,
 865                                                     right, mode);
 866                 ir_node *shift_low    = new_rd_Shr(dbgi, block_true, left_low, right,
 867                                                    low_unsigned);
 868                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 869                                                    low_unsigned);
 870                 ir_node *conv         = create_conv(block_true, left_high,
 871                                                     low_unsigned);
 872                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
 873                 ir_node *carry0       = new_rd_Shl(dbgi, block_true, conv, one,
 874                                                    low_unsigned);
 875                 ir_node *carry1       = new_rd_Shl(dbgi, block_true, carry0,
 876                                                    not_shiftval, low_unsigned);
 877                 ir_node *res_low      = new_rd_Or(dbgi, block_true, shift_low, carry1,
 878                                                   low_unsigned);
 879                 lower_in[0]           = new_r_Jmp(block_true);
 880                 phi_low_in[0]         = res_low;
 881                 phi_high_in[0]        = res_high;
 882         }
 883
 884         /* false block => shift_width > 1word */
 885         {
 886                 ir_node *in[1]       = { proj_false };
 887                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
 888                 ir_node *conv        = create_conv(block_false, left_high, low_unsigned);
 889                 ir_node *res_low     = new_rd_shrs(dbgi, block_false, conv, right,
 890                                                    low_unsigned);
 891                 int      cnsti       = modulo_shift2-1;
 892                 ir_node *cnst2       = new_r_Const_long(irg, low_unsigned, cnsti);
 893                 ir_node *res_high;
 894                 if (new_rd_shrs == new_rd_Shrs) {
 895                         res_high = new_rd_shrs(dbgi, block_false, left_high, cnst2, mode);
 896                 } else {
 897                         res_high = new_r_Const(irg, get_mode_null(mode));
 898                 }
 899                 lower_in[1]          = new_r_Jmp(block_false);
 900                 phi_low_in[1]        = res_low;
 901                 phi_high_in[1]       = res_high;
 902         }
 903
 904         /* patch lower block */
 905         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
 906         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
 907                              low_unsigned);
 908         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
 909                              mode);
 910         ir_set_dw_lowered(node, phi_low, phi_high);
 911 }
 912
 913 static void lower_Shr(ir_node *node, ir_mode *mode)
 914 {
 915         lower_shr_helper(node, mode, new_rd_Shr);
 916 }
 917
 918 static void lower_Shrs(ir_node *node, ir_mode *mode)
 919 {
 920         lower_shr_helper(node, mode, new_rd_Shrs);
 921 }
 922
 923 static void lower_Shl(ir_node *node, ir_mode *mode)
 924 {
 925         ir_node  *right         = get_binop_right(node);
 926         ir_node  *left          = get_binop_left(node);
 927         ir_mode  *shr_mode      = get_irn_mode(node);
 928         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 929         ir_mode  *low_unsigned  = env->low_unsigned;
 930         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 931         ir_graph *irg           = get_irn_irg(node);
 932         ir_node  *left_low      = get_lowered_low(left);
 933         ir_node  *left_high     = get_lowered_high(left);
 934         dbg_info *dbgi          = get_irn_dbg_info(node);
 935         ir_node  *lower_block   = get_nodes_block(node);
 936         ir_node  *block;
 937         ir_node  *cnst;
 938         ir_node  *andn;
 939         ir_node  *cmp;
 940         ir_node  *cond;
 941         ir_node  *proj_true;
 942         ir_node  *proj_false;
 943         ir_node  *phi_low;
 944         ir_node  *phi_high;
 945         ir_node  *lower_in[2];
 946         ir_node  *phi_low_in[2];
 947         ir_node  *phi_high_in[2];
 948
 949         /* this version is optimized for modulo shift architectures
 950          * (and can't handle anything else) */
 951         if (modulo_shift != get_mode_size_bits(shr_mode)
 952                         || modulo_shift2<<1 != modulo_shift) {
 953                 panic("Shl lowering only implemented for modulo shift shl operations");
 954         }
 955         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 956                 panic("Shl lowering only implemented for power-of-2 modes");
 957         }
 958         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 959         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 960                 panic("Shl lowering only implemented for two-complement modes");
 961         }
 962
 963         /* if the right operand is a 64bit value, we're only interested in the
 964          * lower word */
 965         if (get_irn_mode(right) == env->high_unsigned) {
 966                 right = get_lowered_low(right);
 967         } else {
 968                 /* shift should never have signed mode on the right */
 969                 assert(get_irn_mode(right) != env->high_signed);
 970                 right = create_conv(lower_block, right, low_unsigned);
 971         }
 972
 973         part_block_dw(node);
 974         env->flags |= CF_CHANGED;
 975         block = get_nodes_block(node);
 976
 977         /* add a Cmp to test if highest bit is set <=> whether we shift more
 978          * than half the word width */
 979         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 980         andn       = new_r_And(block, right, cnst, low_unsigned);
 981         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 982         cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
 983         cond       = new_rd_Cond(dbgi, block, cmp);
 984         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 985         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 986
 987         /* the true block => shift_width < 1word */
 988         {
 989                 ir_node *in[1]        = { proj_true };
 990                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 991
 992                 ir_node *res_low      = new_rd_Shl(dbgi, block_true, left_low,
 993                                                    right, low_unsigned);
 994                 ir_node *shift_high   = new_rd_Shl(dbgi, block_true, left_high, right,
 995                                                    mode);
 996                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 997                                                    low_unsigned);
 998                 ir_node *conv         = create_conv(block_true, left_low, mode);
 999                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
1000                 ir_node *carry0       = new_rd_Shr(dbgi, block_true, conv, one, mode);
1001                 ir_node *carry1       = new_rd_Shr(dbgi, block_true, carry0,
1002                                                    not_shiftval, mode);
1003                 ir_node *res_high     = new_rd_Or(dbgi, block_true, shift_high, carry1,
1004                                                   mode);
1005                 lower_in[0]           = new_r_Jmp(block_true);
1006                 phi_low_in[0]         = res_low;
1007                 phi_high_in[0]        = res_high;
1008         }
1009
1010         /* false block => shift_width > 1word */
1011         {
1012                 ir_node *in[1]       = { proj_false };
1013                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
1014                 ir_node *res_low     = new_r_Const(irg, get_mode_null(low_unsigned));
1015                 ir_node *conv        = create_conv(block_false, left_low, mode);
1016                 ir_node *res_high    = new_rd_Shl(dbgi, block_false, conv, right, mode);
1017                 lower_in[1]          = new_r_Jmp(block_false);
1018                 phi_low_in[1]        = res_low;
1019                 phi_high_in[1]       = res_high;
1020         }
1021
1022         /* patch lower block */
1023         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
1024         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
1025                              low_unsigned);
1026         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
1027                              mode);
1028         ir_set_dw_lowered(node, phi_low, phi_high);
1029 }
1030
1031 /**
1032  * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
1033  */
1034 static void prepare_links_and_handle_rotl(ir_node *node, void *data)
1035 {
1036         (void) data;
1037         if (is_Rotl(node)) {
1038                 ir_mode  *mode = get_irn_op_mode(node);
1039                 ir_node  *right;
1040                 ir_node  *left, *shl, *shr, *ornode, *block, *sub, *c;
1041                 ir_mode  *omode, *rmode;
1042                 ir_graph *irg;
1043                 dbg_info *dbg;
1044                 optimization_state_t state;
1045
1046                 if (mode != env->high_signed && mode != env->high_unsigned) {
1047                         prepare_links(node);
1048                         return;
1049                 }
1050
1051                 /* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
1052                 right = get_Rotl_right(node);
1053                 irg   = get_irn_irg(node);
1054                 dbg   = get_irn_dbg_info(node);
1055                 omode = get_irn_mode(node);
1056                 left  = get_Rotl_left(node);
1057                 block = get_nodes_block(node);
1058                 shl   = new_rd_Shl(dbg, block, left, right, omode);
1059                 rmode = get_irn_mode(right);
1060                 c     = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
1061                 sub   = new_rd_Sub(dbg, block, c, right, rmode);
1062                 shr   = new_rd_Shr(dbg, block, left, sub, omode);
1063
1064                 /* switch optimization off here, or we will get the Rotl back */
1065                 save_optimization_state(&state);
1066                 set_opt_algebraic_simplification(0);
1067                 ornode = new_rd_Or(dbg, block, shl, shr, omode);
1068                 restore_optimization_state(&state);
1069
1070                 exchange(node, ornode);
1071
1072                 /* do lowering on the new nodes */
1073                 prepare_links(shl);
1074                 prepare_links(c);
1075                 prepare_links(sub);
1076                 prepare_links(shr);
1077                 prepare_links(ornode);
1078                 return;
1079         }
1080
1081         prepare_links(node);
1082 }
1083
1084 /**
1085  * Translate an Unop.
1086  *
1087  * Create an intrinsic Call.
1088  */
1089 static void lower_unop(ir_node *node, ir_mode *mode)
1090 {
1091         ir_node  *op       = get_unop_op(node);
1092         dbg_info *dbgi     = get_irn_dbg_info(node);
1093         ir_node  *block    = get_nodes_block(node);
1094         ir_graph *irg      = get_irn_irg(block);
1095         ir_type  *mtp      = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
1096         ir_op    *irop     = get_irn_op(node);
1097         ir_node  *addr     = get_intrinsic_address(mtp, irop, mode, mode);
1098         ir_node  *nomem    = get_irg_no_mem(irg);
1099         ir_node  *in[2];
1100         ir_node  *call;
1101         ir_node  *resproj;
1102
1103         if (env->params->little_endian) {
1104                 in[0] = get_lowered_low(op);
1105                 in[1] = get_lowered_high(op);
1106         } else {
1107                 in[0] = get_lowered_high(op);
1108                 in[1] = get_lowered_low(op);
1109         }
1110         call    = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
1111         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
1112         set_irn_pinned(call, get_irn_pinned(node));
1113
1114         if (env->params->little_endian) {
1115                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
1116                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
1117                 ir_set_dw_lowered(node, res_low, res_high);
1118         } else {
1119                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
1120                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
1121                 ir_set_dw_lowered(node, res_low, res_high);
1122         }
1123 }
1124
1125 /**
1126  * Translate a logical binop.
1127  *
1128  * Create two logical binops.
1129  */
1130 static void lower_binop_logical(ir_node *node, ir_mode *mode,
1131                                                                 ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
1132 {
1133         ir_node               *left        = get_binop_left(node);
1134         ir_node               *right       = get_binop_right(node);
1135         const lower64_entry_t *left_entry  = get_node_entry(left);
1136         const lower64_entry_t *right_entry = get_node_entry(right);
1137         dbg_info              *dbgi        = get_irn_dbg_info(node);
1138         ir_node               *block       = get_nodes_block(node);
1139         ir_node               *res_low
1140                 = constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
1141                             env->low_unsigned);
1142         ir_node               *res_high
1143                 = constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
1144                             mode);
1145         ir_set_dw_lowered(node, res_low, res_high);
1146 }
1147
1148 static void lower_And(ir_node *node, ir_mode *mode)
1149 {
1150         lower_binop_logical(node, mode, new_rd_And);
1151 }
1152
1153 static void lower_Or(ir_node *node, ir_mode *mode)
1154 {
1155         lower_binop_logical(node, mode, new_rd_Or);
1156 }
1157
1158 static void lower_Eor(ir_node *node, ir_mode *mode)
1159 {
1160         lower_binop_logical(node, mode, new_rd_Eor);
1161 }
1162
1163 /**
1164  * Translate a Not.
1165  *
1166  * Create two logical Nots.
1167  */
1168 static void lower_Not(ir_node *node, ir_mode *mode)
1169 {
1170         ir_node               *op       = get_Not_op(node);
1171         const lower64_entry_t *op_entry = get_node_entry(op);
1172         dbg_info              *dbgi     = get_irn_dbg_info(node);
1173         ir_node               *block    = get_nodes_block(node);
1174         ir_node               *res_low
1175                 = new_rd_Not(dbgi, block, op_entry->low_word, env->low_unsigned);
1176         ir_node               *res_high
1177                 = new_rd_Not(dbgi, block, op_entry->high_word, mode);
1178         ir_set_dw_lowered(node, res_low, res_high);
1179 }
1180
1181 static void lower_Proj(ir_node *node, ir_mode *op_mode)
1182 {
1183         ir_mode *mode = get_irn_mode(node);
1184         ir_node *pred;
1185         (void)op_mode;
1186         if (mode != env->high_signed && mode != env->high_unsigned)
1187                 return;
1188         /* skip tuples */
1189         pred = get_Proj_pred(node);
1190         if (is_Tuple(pred)) {
1191                 long                   pn    = get_Proj_proj(node);
1192                 ir_node               *op    = get_irn_n(pred, pn);
1193                 const lower64_entry_t *entry = get_node_entry(op);
1194                 ir_set_dw_lowered(node, entry->low_word, entry->high_word);
1195         }
1196 }
1197
1198 static bool is_equality_cmp(const ir_node *node)
1199 {
1200         ir_relation relation = get_Cmp_relation(node);
1201         ir_node    *left     = get_Cmp_left(node);
1202         ir_node    *right    = get_Cmp_right(node);
1203         ir_mode    *mode     = get_irn_mode(left);
1204
1205         /* this probably makes no sense if unordered is involved */
1206         assert(!mode_is_float(mode));
1207
1208         if (relation == ir_relation_equal || relation == ir_relation_less_greater)
1209                 return true;
1210
1211         if (!is_Const(right) || !is_Const_null(right))
1212                 return false;
1213         if (mode_is_signed(mode)) {
1214                 return relation == ir_relation_less_greater;
1215         } else {
1216                 return relation == ir_relation_greater;
1217         }
1218 }
1219
1220 static ir_node *get_cfop_destination(const ir_node *cfop)
1221 {
1222         const ir_edge_t *first = get_irn_out_edge_first(cfop);
1223         /* we should only have 1 destination */
1224         assert(get_irn_n_edges(cfop) == 1);
1225         return get_edge_src_irn(first);
1226 }
1227
1228 static void lower_Switch(ir_node *node, ir_mode *high_mode)
1229 {
1230         ir_node *selector = get_Switch_selector(node);
1231         ir_mode *mode     = get_irn_mode(selector);
1232         (void)high_mode;
1233         if (mode == env->high_signed || mode == env->high_unsigned) {
1234                 /* we can't really handle Switch with 64bit offsets */
1235                 panic("Switch with 64bit jumptable not supported");
1236         }
1237         lower_node(selector);
1238 }
1239
1240 /**
1241  * Translate a Cond.
1242  */
1243 static void lower_Cond(ir_node *node, ir_mode *high_mode)
1244 {
1245         ir_node *left, *right, *block;
1246         ir_node *sel = get_Cond_selector(node);
1247         ir_mode *cmp_mode;
1248         const lower64_entry_t *lentry, *rentry;
1249         ir_node  *projT = NULL, *projF = NULL;
1250         ir_node  *new_bl, *irn;
1251         ir_node  *projHF, *projHT;
1252         ir_node  *dst_blk;
1253         ir_relation relation;
1254         ir_graph *irg;
1255         dbg_info *dbg;
1256
1257         (void) high_mode;
1258
1259         if (!is_Cmp(sel)) {
1260                 lower_node(sel);
1261                 return;
1262         }
1263
1264         left     = get_Cmp_left(sel);
1265         cmp_mode = get_irn_mode(left);
1266         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1267                 lower_node(sel);
1268                 return;
1269         }
1270
1271         right  = get_Cmp_right(sel);
1272         lower_node(left);
1273         lower_node(right);
1274         lentry = get_node_entry(left);
1275         rentry = get_node_entry(right);
1276
1277         /* all right, build the code */
1278         foreach_out_edge_safe(node, edge) {
1279                 ir_node *proj    = get_edge_src_irn(edge);
1280                 long     proj_nr;
1281                 if (!is_Proj(proj))
1282                         continue;
1283                 proj_nr = get_Proj_proj(proj);
1284
1285                 if (proj_nr == pn_Cond_true) {
1286                         assert(projT == NULL && "more than one Proj(true)");
1287                         projT = proj;
1288                 } else {
1289                         assert(proj_nr == pn_Cond_false);
1290                         assert(projF == NULL && "more than one Proj(false)");
1291                         projF = proj;
1292                 }
1293                 mark_irn_visited(proj);
1294         }
1295         assert(projT && projF);
1296
1297         /* create a new high compare */
1298         block    = get_nodes_block(node);
1299         irg      = get_Block_irg(block);
1300         dbg      = get_irn_dbg_info(sel);
1301         relation = get_Cmp_relation(sel);
1302
1303         if (is_equality_cmp(sel)) {
1304                 /* x ==/!= y ==> or(x_low^y_low,x_high^y_high) ==/!= 0 */
1305                 ir_mode *mode       = env->low_unsigned;
1306                 ir_node *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1307                 ir_node *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1308                 ir_node *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1309                 ir_node *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1310                 ir_node *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1311                 ir_node *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1312                 ir_node *ornode = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1313                 ir_node *cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1314                 set_Cond_selector(node, cmp);
1315                 return;
1316         }
1317
1318         if (relation == ir_relation_equal) {
1319                 ir_node *proj;
1320                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1321                 dst_blk = get_cfop_destination(projF);
1322
1323                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1324                                  ir_relation_equal);
1325                 dbg = get_irn_dbg_info(node);
1326                 irn = new_rd_Cond(dbg, block, irn);
1327
1328                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1329                 mark_irn_visited(projHF);
1330                 exchange(projF, projHF);
1331
1332                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1333                 mark_irn_visited(projHT);
1334
1335                 new_bl = new_r_Block(irg, 1, &projHT);
1336
1337                 dbg = get_irn_dbg_info(sel);
1338                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1339                                   ir_relation_equal);
1340                 dbg = get_irn_dbg_info(node);
1341                 irn = new_rd_Cond(dbg, new_bl, irn);
1342
1343                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1344                 mark_irn_visited(proj);
1345                 add_block_cf_input(dst_blk, projHF, proj);
1346
1347                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1348                 mark_irn_visited(proj);
1349                 exchange(projT, proj);
1350         } else if (relation == ir_relation_less_greater) {
1351                 ir_node *proj;
1352                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1353                 dst_blk = get_cfop_destination(projT);
1354
1355                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1356                                  ir_relation_less_greater);
1357                 dbg = get_irn_dbg_info(node);
1358                 irn = new_rd_Cond(dbg, block, irn);
1359
1360                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1361                 mark_irn_visited(projHT);
1362                 exchange(projT, projHT);
1363
1364                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1365                 mark_irn_visited(projHF);
1366
1367                 new_bl = new_r_Block(irg, 1, &projHF);
1368
1369                 dbg = get_irn_dbg_info(sel);
1370                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1371                                  ir_relation_less_greater);
1372                 dbg = get_irn_dbg_info(node);
1373                 irn = new_rd_Cond(dbg, new_bl, irn);
1374
1375                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1376                 mark_irn_visited(proj);
1377                 add_block_cf_input(dst_blk, projHT, proj);
1378
1379                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1380                 mark_irn_visited(proj);
1381                 exchange(projF, proj);
1382         } else {
1383                 ir_node *proj;
1384                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1385                 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1386                 ir_node *projEqF;
1387
1388                 dstT = get_cfop_destination(projT);
1389                 dstF = get_cfop_destination(projF);
1390
1391                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1392                                  relation & ~ir_relation_equal);
1393                 dbg = get_irn_dbg_info(node);
1394                 irn = new_rd_Cond(dbg, block, irn);
1395
1396                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1397                 mark_irn_visited(projHT);
1398
1399                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1400                 mark_irn_visited(projHF);
1401
1402                 newbl_eq = new_r_Block(irg, 1, &projHF);
1403
1404                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1405                                  ir_relation_equal);
1406                 irn = new_rd_Cond(dbg, newbl_eq, irn);
1407
1408                 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1409                 mark_irn_visited(projEqF);
1410
1411                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1412                 mark_irn_visited(proj);
1413
1414                 newbl_l = new_r_Block(irg, 1, &proj);
1415
1416                 dbg = get_irn_dbg_info(sel);
1417                 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1418                                  relation);
1419                 dbg = get_irn_dbg_info(node);
1420                 irn = new_rd_Cond(dbg, newbl_l, irn);
1421
1422                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1423                 mark_irn_visited(proj);
1424                 add_block_cf_input(dstT, projT, proj);
1425
1426                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1427                 mark_irn_visited(proj);
1428                 add_block_cf_input(dstF, projF, proj);
1429
1430                 exchange(projT, projHT);
1431                 exchange(projF, projEqF);
1432         }
1433
1434         /* we have changed the control flow */
1435         env->flags |= CF_CHANGED;
1436 }
1437
1438 /**
1439  * Translate a Conv to higher_signed
1440  */
1441 static void lower_Conv_to_Ll(ir_node *node)
1442 {
1443         ir_mode  *omode        = get_irn_mode(node);
1444         ir_node  *op           = get_Conv_op(node);
1445         ir_mode  *imode        = get_irn_mode(op);
1446         ir_graph *irg          = get_irn_irg(node);
1447         ir_node  *block        = get_nodes_block(node);
1448         dbg_info *dbg          = get_irn_dbg_info(node);
1449         ir_node  *res_low;
1450         ir_node  *res_high;
1451
1452         ir_mode  *low_unsigned = env->low_unsigned;
1453         ir_mode  *low_signed
1454                 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1455
1456         if (mode_is_int(imode) || mode_is_reference(imode)) {
1457                 if (imode == env->high_signed || imode == env->high_unsigned) {
1458                         /* a Conv from Lu to Ls or Ls to Lu */
1459                         const lower64_entry_t *op_entry = get_node_entry(op);
1460                         res_low  = op_entry->low_word;
1461                         res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1462                 } else {
1463                         /* simple case: create a high word */
1464                         if (imode != low_unsigned)
1465                                 op = new_rd_Conv(dbg, block, op, low_unsigned);
1466
1467                         res_low = op;
1468
1469                         if (mode_is_signed(imode)) {
1470                                 int      c       = get_mode_size_bits(low_signed) - 1;
1471                                 ir_node *cnst    = new_r_Const_long(irg, low_unsigned, c);
1472                                 if (get_irn_mode(op) != low_signed)
1473                                         op = new_rd_Conv(dbg, block, op, low_signed);
1474                                 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1475                         } else {
1476                                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1477                         }
1478                 }
1479         } else if (imode == mode_b) {
1480                 res_low  = new_rd_Conv(dbg, block, op, low_unsigned);
1481                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1482         } else {
1483                 ir_node *irn, *call;
1484                 ir_type *mtp = get_conv_type(imode, omode);
1485
1486                 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1487                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1488                 set_irn_pinned(call, get_irn_pinned(node));
1489                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1490
1491                 if (env->params->little_endian) {
1492                         res_low  = new_r_Proj(irn, low_unsigned, 0);
1493                         res_high = new_r_Proj(irn, low_signed, 1);
1494                 } else {
1495                         res_low  = new_r_Proj(irn, low_unsigned, 1);
1496                         res_high = new_r_Proj(irn, low_signed,   0);
1497                 }
1498         }
1499         ir_set_dw_lowered(node, res_low, res_high);
1500 }
1501
1502 /**
1503  * Translate a Conv from higher_unsigned
1504  */
1505 static void lower_Conv_from_Ll(ir_node *node)
1506 {
1507         ir_node               *op    = get_Conv_op(node);
1508         ir_mode               *omode = get_irn_mode(node);
1509         ir_node               *block = get_nodes_block(node);
1510         dbg_info              *dbg   = get_irn_dbg_info(node);
1511         ir_graph              *irg   = get_irn_irg(node);
1512         const lower64_entry_t *entry = get_node_entry(op);
1513
1514         if (mode_is_int(omode) || mode_is_reference(omode)) {
1515                 op = entry->low_word;
1516
1517                 /* simple case: create a high word */
1518                 if (omode != env->low_unsigned)
1519                         op = new_rd_Conv(dbg, block, op, omode);
1520
1521                 set_Conv_op(node, op);
1522         } else if (omode == mode_b) {
1523                 /* llu ? true : false  <=> (low|high) ? true : false */
1524                 ir_mode *mode   = env->low_unsigned;
1525                 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1526                                             entry->high_word, mode);
1527                 set_Conv_op(node, ornode);
1528         } else {
1529                 ir_node *irn, *call, *in[2];
1530                 ir_mode *imode = get_irn_mode(op);
1531                 ir_type *mtp   = get_conv_type(imode, omode);
1532                 ir_node *res;
1533
1534                 irn   = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1535                 if (env->params->little_endian) {
1536                         in[0] = entry->low_word;
1537                         in[1] = entry->high_word;
1538                 } else {
1539                         in[0] = entry->high_word;
1540                         in[1] = entry->low_word;
1541                 }
1542
1543                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1544                 set_irn_pinned(call, get_irn_pinned(node));
1545                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1546                 res = new_r_Proj(irn, omode, 0);
1547
1548                 exchange(node, res);
1549         }
1550 }
1551
1552 /**
1553  * lower Cmp
1554  */
1555 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1556 {
1557         ir_node  *l        = get_Cmp_left(cmp);
1558         ir_mode  *cmp_mode = get_irn_mode(l);
1559         ir_node  *r, *low, *high, *t, *res;
1560         ir_relation relation;
1561         ir_node  *block;
1562         dbg_info *dbg;
1563         const lower64_entry_t *lentry;
1564         const lower64_entry_t *rentry;
1565         (void) m;
1566
1567         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned)
1568                 return;
1569
1570         r        = get_Cmp_right(cmp);
1571         lentry   = get_node_entry(l);
1572         rentry   = get_node_entry(r);
1573         relation = get_Cmp_relation(cmp);
1574         block    = get_nodes_block(cmp);
1575         dbg      = get_irn_dbg_info(cmp);
1576
1577         /* easy case for x ==/!= 0 (see lower_Cond for details) */
1578         if (is_equality_cmp(cmp)) {
1579                 ir_graph *irg        = get_irn_irg(cmp);
1580                 ir_mode  *mode       = env->low_unsigned;
1581                 ir_node  *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1582                 ir_node  *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1583                 ir_node  *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1584                 ir_node  *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1585                 ir_node  *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1586                 ir_node  *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1587                 ir_node  *ornode     = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1588                 ir_node  *new_cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1589                 exchange(cmp, new_cmp);
1590                 return;
1591         }
1592
1593         if (relation == ir_relation_equal) {
1594                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1595                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1596                                   relation);
1597                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1598                                   relation);
1599                 res  = new_rd_And(dbg, block, low, high, mode_b);
1600         } else if (relation == ir_relation_less_greater) {
1601                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1602                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1603                                   relation);
1604                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1605                                   relation);
1606                 res = new_rd_Or(dbg, block, low, high, mode_b);
1607         } else {
1608                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1609                 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1610                         rentry->high_word, relation & ~ir_relation_equal);
1611                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1612                                   relation);
1613                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1614                                   ir_relation_equal);
1615                 t = new_rd_And(dbg, block, low, high, mode_b);
1616                 res = new_rd_Or(dbg, block, high1, t, mode_b);
1617         }
1618         exchange(cmp, res);
1619 }
1620
1621 /**
1622  * Translate a Conv.
1623  */
1624 static void lower_Conv(ir_node *node, ir_mode *mode)
1625 {
1626         mode = get_irn_mode(node);
1627
1628         if (mode == env->high_signed || mode == env->high_unsigned) {
1629                 lower_Conv_to_Ll(node);
1630         } else {
1631                 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1632
1633                 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1634                         lower_Conv_from_Ll(node);
1635                 }
1636         }
1637 }
1638
1639 static void fix_parameter_entities(ir_graph *irg, ir_type *orig_mtp)
1640 {
1641         size_t      orig_n_params      = get_method_n_params(orig_mtp);
1642         ir_entity **parameter_entities;
1643
1644         parameter_entities = ALLOCANZ(ir_entity*, orig_n_params);
1645
1646         ir_type *frame_type = get_irg_frame_type(irg);
1647         size_t   n          = get_compound_n_members(frame_type);
1648         size_t   i;
1649         size_t   n_param;
1650
1651         /* collect parameter entities */
1652         for (i = 0; i < n; ++i) {
1653                 ir_entity *entity = get_compound_member(frame_type, i);
1654                 size_t     p;
1655                 if (!is_parameter_entity(entity))
1656                         continue;
1657                 p = get_entity_parameter_number(entity);
1658                 if (p == IR_VA_START_PARAMETER_NUMBER)
1659                         continue;
1660                 assert(p < orig_n_params);
1661                 assert(parameter_entities[p] == NULL);
1662                 parameter_entities[p] = entity;
1663         }
1664
1665         /* adjust indices */
1666         n_param = 0;
1667         for (i = 0; i < orig_n_params; ++i, ++n_param) {
1668                 ir_entity *entity = parameter_entities[i];
1669                 ir_type   *tp;
1670
1671                 if (entity != NULL)
1672                         set_entity_parameter_number(entity, n_param);
1673
1674                 tp = get_method_param_type(orig_mtp, i);
1675                 if (is_Primitive_type(tp)) {
1676                         ir_mode *mode = get_type_mode(tp);
1677                         if (mode == env->high_signed || mode == env->high_unsigned) {
1678                                 ++n_param;
1679                                 /* note that we do not change the type of the parameter
1680                                  * entities, as calling convention fixup later still needs to
1681                                  * know which is/was a lowered doubleword.
1682                                  * So we just mark/remember it for later */
1683                                 if (entity != NULL) {
1684                                         assert(entity->attr.parameter.doubleword_low_mode == NULL);
1685                                         entity->attr.parameter.doubleword_low_mode
1686                                                 = env->low_unsigned;
1687                                 }
1688                         }
1689                 }
1690         }
1691 }
1692
1693 /**
1694  * Lower the method type.
1695  *
1696  * @param env  the lower environment
1697  * @param mtp  the method type to lower
1698  *
1699  * @return the lowered type
1700  */
1701 static ir_type *lower_mtp(ir_type *mtp)
1702 {
1703         ir_type *res;
1704         size_t   i;
1705         size_t   orig_n_params;
1706         size_t   orig_n_res;
1707         size_t   n_param;
1708         size_t   n_res;
1709         bool     must_be_lowered;
1710
1711         res = pmap_get(ir_type, lowered_type, mtp);
1712         if (res != NULL)
1713                 return res;
1714         if (type_visited(mtp))
1715                 return mtp;
1716         mark_type_visited(mtp);
1717
1718         orig_n_params   = get_method_n_params(mtp);
1719         orig_n_res      = get_method_n_ress(mtp);
1720         n_param         = orig_n_params;
1721         n_res           = orig_n_res;
1722         must_be_lowered = false;
1723
1724         /* count new number of params */
1725         for (i = orig_n_params; i > 0;) {
1726                 ir_type *tp = get_method_param_type(mtp, --i);
1727
1728                 if (is_Primitive_type(tp)) {
1729                         ir_mode *mode = get_type_mode(tp);
1730
1731                         if (mode == env->high_signed || mode == env->high_unsigned) {
1732                                 ++n_param;
1733                                 must_be_lowered = true;
1734                         }
1735                 }
1736         }
1737
1738         /* count new number of results */
1739         for (i = orig_n_res; i > 0;) {
1740                 ir_type *tp = get_method_res_type(mtp, --i);
1741
1742                 if (is_Primitive_type(tp)) {
1743                         ir_mode *mode = get_type_mode(tp);
1744
1745                         if (mode == env->high_signed || mode == env->high_unsigned) {
1746                                 ++n_res;
1747                                 must_be_lowered = true;
1748                         }
1749                 }
1750         }
1751         if (!must_be_lowered) {
1752                 set_type_link(mtp, NULL);
1753                 return mtp;
1754         }
1755
1756         res = new_d_type_method(n_param, n_res, get_type_dbg_info(mtp));
1757
1758         /* set param types and result types */
1759         for (i = n_param = 0; i < orig_n_params; ++i) {
1760                 ir_type *tp = get_method_param_type(mtp, i);
1761
1762                 if (is_Primitive_type(tp)) {
1763                         ir_mode *mode = get_type_mode(tp);
1764
1765                         if (mode == env->high_signed) {
1766                                 if (env->params->little_endian) {
1767                                         set_method_param_type(res, n_param++, tp_u);
1768                                         set_method_param_type(res, n_param++, tp_s);
1769                                 } else {
1770                                         set_method_param_type(res, n_param++, tp_s);
1771                                         set_method_param_type(res, n_param++, tp_u);
1772                                 }
1773                         } else if (mode == env->high_unsigned) {
1774                                 set_method_param_type(res, n_param++, tp_u);
1775                                 set_method_param_type(res, n_param++, tp_u);
1776                         } else {
1777                                 set_method_param_type(res, n_param, tp);
1778                                 ++n_param;
1779                         }
1780                 } else {
1781                         set_method_param_type(res, n_param, tp);
1782                         ++n_param;
1783                 }
1784         }
1785         for (i = n_res = 0; i < orig_n_res; ++i) {
1786                 ir_type *tp = get_method_res_type(mtp, i);
1787
1788                 if (is_Primitive_type(tp)) {
1789                         ir_mode *mode = get_type_mode(tp);
1790
1791                         if (mode == env->high_signed) {
1792                                 if (env->params->little_endian) {
1793                                         set_method_res_type(res, n_res++, tp_u);
1794                                         set_method_res_type(res, n_res++, tp_s);
1795                                 } else {
1796                                         set_method_res_type(res, n_res++, tp_s);
1797                                         set_method_res_type(res, n_res++, tp_u);
1798                                 }
1799                         } else if (mode == env->high_unsigned) {
1800                                 set_method_res_type(res, n_res++, tp_u);
1801                                 set_method_res_type(res, n_res++, tp_u);
1802                         } else {
1803                                 set_method_res_type(res, n_res++, tp);
1804                         }
1805                 } else {
1806                         set_method_res_type(res, n_res++, tp);
1807                 }
1808         }
1809
1810         set_method_variadicity(res, get_method_variadicity(mtp));
1811         set_method_calling_convention(res, get_method_calling_convention(mtp));
1812         set_method_additional_properties(res, get_method_additional_properties(mtp));
1813
1814         set_higher_type(res, mtp);
1815         set_type_link(res, mtp);
1816
1817         mark_type_visited(res);
1818         pmap_insert(lowered_type, mtp, res);
1819         return res;
1820 }
1821
1822 /**
1823  * Translate a Return.
1824  */
1825 static void lower_Return(ir_node *node, ir_mode *mode)
1826 {
1827         ir_node  **in;
1828         size_t     i, j, n;
1829         int        need_conv = 0;
1830         (void) mode;
1831
1832         /* check if this return must be lowered */
1833         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1834                 ir_node *pred  = get_Return_res(node, i);
1835                 ir_mode *rmode = get_irn_op_mode(pred);
1836
1837                 if (rmode == env->high_signed || rmode == env->high_unsigned)
1838                         need_conv = 1;
1839         }
1840         if (! need_conv)
1841                 return;
1842
1843         ir_graph  *irg = get_irn_irg(node);
1844         ir_entity *ent = get_irg_entity(irg);
1845         ir_type   *mtp = get_entity_type(ent);
1846
1847         /* create a new in array */
1848         NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1849         j = 0;
1850         in[j++] = get_Return_mem(node);
1851
1852         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1853                 ir_node *pred      = get_Return_res(node, i);
1854                 ir_mode *pred_mode = get_irn_mode(pred);
1855
1856                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1857                         const lower64_entry_t *entry = get_node_entry(pred);
1858                         if (env->params->little_endian) {
1859                                 in[j++] = entry->low_word;
1860                                 in[j++] = entry->high_word;
1861                         } else {
1862                                 in[j++] = entry->high_word;
1863                                 in[j++] = entry->low_word;
1864                         }
1865                 } else {
1866                         in[j++] = pred;
1867                 }
1868         }
1869         assert(j == get_method_n_ress(mtp)+1);
1870
1871         set_irn_in(node, j, in);
1872 }
1873
1874 /**
1875  * Translate the parameters.
1876  */
1877 static void lower_Start(ir_node *node, ir_mode *high_mode)
1878 {
1879         ir_graph  *irg      = get_irn_irg(node);
1880         ir_entity *ent      = get_irg_entity(irg);
1881         ir_type   *mtp      = get_entity_type(ent);
1882         ir_type   *orig_mtp = (ir_type*)get_type_link(mtp);
1883         ir_node   *args;
1884         long      *new_projs;
1885         size_t    i, j, n_params;
1886         (void) high_mode;
1887
1888         /* if type link is NULL then the type was not lowered, hence no changes
1889          * at Start necessary */
1890         if (orig_mtp == NULL)
1891                 return;
1892
1893         n_params = get_method_n_params(orig_mtp);
1894
1895         NEW_ARR_A(long, new_projs, n_params);
1896
1897         /* Calculate mapping of proj numbers in new_projs */
1898         for (i = j = 0; i < n_params; ++i, ++j) {
1899                 ir_type *ptp = get_method_param_type(orig_mtp, i);
1900
1901                 new_projs[i] = j;
1902                 if (is_Primitive_type(ptp)) {
1903                         ir_mode *amode = get_type_mode(ptp);
1904                         if (amode == env->high_signed || amode == env->high_unsigned)
1905                                 ++j;
1906                 }
1907         }
1908
1909         /* find args Proj */
1910         args = NULL;
1911         foreach_out_edge(node, edge) {
1912                 ir_node *proj = get_edge_src_irn(edge);
1913                 if (!is_Proj(proj))
1914                         continue;
1915                 if (get_Proj_proj(proj) == pn_Start_T_args) {
1916                         args = proj;
1917                         break;
1918                 }
1919         }
1920         if (args == NULL)
1921                 return;
1922
1923         /* fix all Proj's and create new ones */
1924         foreach_out_edge_safe(args, edge) {
1925                 ir_node *proj   = get_edge_src_irn(edge);
1926                 ir_mode *mode   = get_irn_mode(proj);
1927                 ir_mode *mode_l = env->low_unsigned;
1928                 ir_node *pred;
1929                 long     proj_nr;
1930                 ir_mode *mode_h;
1931                 ir_node *res_low;
1932                 ir_node *res_high;
1933                 int      old_cse;
1934                 dbg_info *dbg;
1935
1936                 if (!is_Proj(proj))
1937                         continue;
1938                 pred    = get_Proj_pred(proj);
1939                 proj_nr = get_Proj_proj(proj);
1940
1941                 if (mode == env->high_signed) {
1942                         mode_h = env->low_signed;
1943                 } else if (mode == env->high_unsigned) {
1944                         mode_h = env->low_unsigned;
1945                 } else {
1946                         long new_pn = new_projs[proj_nr];
1947                         set_Proj_proj(proj, new_pn);
1948                         continue;
1949                 }
1950
1951                 /* Switch off CSE or we might get an already existing Proj. */
1952                 old_cse = get_opt_cse();
1953                 set_opt_cse(0);
1954                 dbg = get_irn_dbg_info(proj);
1955                 if (env->params->little_endian) {
1956                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1957                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1958                 } else {
1959                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1960                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1961                 }
1962                 set_opt_cse(old_cse);
1963                 ir_set_dw_lowered(proj, res_low, res_high);
1964         }
1965 }
1966
1967 /**
1968  * Translate a Call.
1969  */
1970 static void lower_Call(ir_node *node, ir_mode *mode)
1971 {
1972         ir_type  *tp = get_Call_type(node);
1973         ir_node  **in;
1974         size_t   n_params, n_res;
1975         bool     need_lower = false;
1976         size_t   i, j;
1977         size_t   p;
1978         long     *res_numbers = NULL;
1979         ir_node  *resproj;
1980         (void) mode;
1981
1982         n_params = get_method_n_params(tp);
1983         for (p = 0; p < n_params; ++p) {
1984                 ir_type *ptp = get_method_param_type(tp, p);
1985
1986                 if (is_Primitive_type(ptp)) {
1987                         ir_mode *pmode = get_type_mode(ptp);
1988                         if (pmode == env->high_signed || pmode == env->high_unsigned) {
1989                                 need_lower = true;
1990                                 break;
1991                         }
1992                 }
1993         }
1994         n_res = get_method_n_ress(tp);
1995         if (n_res > 0) {
1996                 NEW_ARR_A(long, res_numbers, n_res);
1997
1998                 for (i = j = 0; i < n_res; ++i, ++j) {
1999                         ir_type *ptp = get_method_res_type(tp, i);
2000
2001                         res_numbers[i] = j;
2002                         if (is_Primitive_type(ptp)) {
2003                                 ir_mode *rmode = get_type_mode(ptp);
2004                                 if (rmode == env->high_signed || rmode == env->high_unsigned) {
2005                                         need_lower = true;
2006                                         ++j;
2007                                 }
2008                         }
2009                 }
2010         }
2011
2012         if (! need_lower)
2013                 return;
2014
2015         /* let's lower it */
2016         tp = lower_mtp(tp);
2017         set_Call_type(node, tp);
2018
2019         NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
2020
2021         in[0] = get_Call_mem(node);
2022         in[1] = get_Call_ptr(node);
2023
2024         for (j = 2, i = 0; i < n_params; ++i) {
2025                 ir_node *pred      = get_Call_param(node, i);
2026                 ir_mode *pred_mode = get_irn_mode(pred);
2027
2028                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
2029                         const lower64_entry_t *pred_entry = get_node_entry(pred);
2030                         if (env->params->little_endian) {
2031                                 in[j++] = pred_entry->low_word;
2032                                 in[j++] = pred_entry->high_word;
2033                         } else {
2034                                 in[j++] = pred_entry->high_word;
2035                                 in[j++] = pred_entry->low_word;
2036                         }
2037                 } else {
2038                         in[j++] = pred;
2039                 }
2040         }
2041
2042         set_irn_in(node, j, in);
2043
2044         /* find results T */
2045         resproj = NULL;
2046         foreach_out_edge(node, edge) {
2047                 ir_node *proj = get_edge_src_irn(edge);
2048                 if (!is_Proj(proj))
2049                         continue;
2050                 if (get_Proj_proj(proj) == pn_Call_T_result) {
2051                         resproj = proj;
2052                         break;
2053                 }
2054         }
2055         if (resproj == NULL)
2056                 return;
2057
2058         /* fix the results */
2059         foreach_out_edge_safe(resproj, edge) {
2060                 ir_node  *proj      = get_edge_src_irn(edge);
2061                 ir_mode  *proj_mode = get_irn_mode(proj);
2062                 ir_mode  *mode_l    = env->low_unsigned;
2063                 ir_node  *pred;
2064                 long      proj_nr;
2065                 ir_mode  *mode_h;
2066                 ir_node  *res_low;
2067                 ir_node  *res_high;
2068                 dbg_info *dbg;
2069
2070                 if (!is_Proj(proj))
2071                         continue;
2072                 pred    = get_Proj_pred(proj);
2073                 proj_nr = get_Proj_proj(proj);
2074
2075                 if (proj_mode == env->high_signed) {
2076                         mode_h = env->low_signed;
2077                 } else if (proj_mode == env->high_unsigned) {
2078                         mode_h = env->low_unsigned;
2079                 } else {
2080                         long new_nr = res_numbers[proj_nr];
2081                         set_Proj_proj(proj, new_nr);
2082                         continue;
2083                 }
2084
2085                 dbg = get_irn_dbg_info(proj);
2086                 if (env->params->little_endian) {
2087                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2088                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2089                 } else {
2090                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2091                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2092                 }
2093                 ir_set_dw_lowered(proj, res_low, res_high);
2094         }
2095 }
2096
2097 /**
2098  * Translate an Unknown into two.
2099  */
2100 static void lower_Unknown(ir_node *node, ir_mode *mode)
2101 {
2102         ir_mode  *low_mode = env->low_unsigned;
2103         ir_graph *irg      = get_irn_irg(node);
2104         ir_node  *res_low  = new_r_Unknown(irg, low_mode);
2105         ir_node  *res_high = new_r_Unknown(irg, mode);
2106         ir_set_dw_lowered(node, res_low, res_high);
2107 }
2108
2109 /**
2110  * Translate a Bad into two.
2111  */
2112 static void lower_Bad(ir_node *node, ir_mode *mode)
2113 {
2114         ir_mode  *low_mode = env->low_unsigned;
2115         ir_graph *irg      = get_irn_irg(node);
2116         ir_node  *res_low  = new_r_Bad(irg, low_mode);
2117         ir_node  *res_high = new_r_Bad(irg, mode);
2118         ir_set_dw_lowered(node, res_low, res_high);
2119 }
2120
2121 /**
2122  * Translate a Phi.
2123  *
2124  * First step: just create two templates
2125  */
2126 static void lower_Phi(ir_node *phi)
2127 {
2128         ir_mode  *mode = get_irn_mode(phi);
2129         int       i;
2130         int       arity;
2131         ir_node **in_l;
2132         ir_node **in_h;
2133         ir_node  *unk_l;
2134         ir_node  *unk_h;
2135         ir_node  *phi_l;
2136         ir_node  *phi_h;
2137         dbg_info *dbg;
2138         ir_node  *block;
2139         ir_graph *irg;
2140         ir_mode  *mode_l;
2141         ir_mode  *mode_h;
2142
2143         /* enqueue predecessors */
2144         arity = get_Phi_n_preds(phi);
2145         for (i = 0; i < arity; ++i) {
2146                 ir_node *pred = get_Phi_pred(phi, i);
2147                 pdeq_putr(env->waitq, pred);
2148         }
2149
2150         if (mode != env->high_signed && mode != env->high_unsigned)
2151                 return;
2152
2153         /* first create a new in array */
2154         NEW_ARR_A(ir_node *, in_l, arity);
2155         NEW_ARR_A(ir_node *, in_h, arity);
2156         irg    = get_irn_irg(phi);
2157         mode_l = env->low_unsigned;
2158         mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2159         unk_l  = new_r_Dummy(irg, mode_l);
2160         unk_h  = new_r_Dummy(irg, mode_h);
2161         for (i = 0; i < arity; ++i) {
2162                 in_l[i] = unk_l;
2163                 in_h[i] = unk_h;
2164         }
2165
2166         dbg   = get_irn_dbg_info(phi);
2167         block = get_nodes_block(phi);
2168         phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2169         phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2170
2171         ir_set_dw_lowered(phi, phi_l, phi_h);
2172
2173         /* remember that we need to fixup the predecessors later */
2174         ARR_APP1(ir_node*, env->lowered_phis, phi);
2175 }
2176
2177 static void fixup_phi(ir_node *phi)
2178 {
2179         const lower64_entry_t *entry = get_node_entry(phi);
2180         ir_node               *phi_l = entry->low_word;
2181         ir_node               *phi_h = entry->high_word;
2182         int                    arity = get_Phi_n_preds(phi);
2183         int                    i;
2184
2185         /* exchange phi predecessors which are lowered by now */
2186         for (i = 0; i < arity; ++i) {
2187                 ir_node               *pred       = get_Phi_pred(phi, i);
2188                 const lower64_entry_t *pred_entry = get_node_entry(pred);
2189
2190                 set_Phi_pred(phi_l, i, pred_entry->low_word);
2191                 set_Phi_pred(phi_h, i, pred_entry->high_word);
2192         }
2193 }
2194
2195 /**
2196  * Translate a Mux.
2197  */
2198 static void lower_Mux(ir_node *mux, ir_mode *mode)
2199 {
2200         ir_node               *truen       = get_Mux_true(mux);
2201         ir_node               *falsen      = get_Mux_false(mux);
2202         ir_node               *sel         = get_Mux_sel(mux);
2203         const lower64_entry_t *true_entry  = get_node_entry(truen);
2204         const lower64_entry_t *false_entry = get_node_entry(falsen);
2205         ir_node               *true_l      = true_entry->low_word;
2206         ir_node               *true_h      = true_entry->high_word;
2207         ir_node               *false_l     = false_entry->low_word;
2208         ir_node               *false_h     = false_entry->high_word;
2209         dbg_info              *dbgi        = get_irn_dbg_info(mux);
2210         ir_node               *block       = get_nodes_block(mux);
2211         ir_node               *res_low
2212                 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2213         ir_node               *res_high
2214                 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2215         ir_set_dw_lowered(mux, res_low, res_high);
2216 }
2217
2218 /**
2219  * Translate an ASM node.
2220  */
2221 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2222 {
2223         ir_mode           *high_signed        = env->high_signed;
2224         ir_mode           *high_unsigned      = env->high_unsigned;
2225         int                n_outs             = get_ASM_n_output_constraints(asmn);
2226         ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2227         ir_asm_constraint *input_constraints  = get_ASM_input_constraints(asmn);
2228         unsigned           n_64bit_outs       = 0;
2229
2230         (void)mode;
2231
2232         for (int i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2233                 ir_node *op      = get_irn_n(asmn, i);
2234                 ir_mode *op_mode = get_irn_mode(op);
2235                 if (op_mode == high_signed || op_mode == high_unsigned) {
2236                         panic("lowering ASM 64bit input unimplemented");
2237                 }
2238         }
2239
2240         for (int o = 0; o < n_outs; ++o) {
2241                 const ir_asm_constraint *constraint = &output_constraints[o];
2242                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2243                         const char *constr = get_id_str(constraint->constraint);
2244                         ++n_64bit_outs;
2245                         /* TODO: How to do this architecture neutral? This is very
2246                          * i386 specific... */
2247                         if (constr[0] != '=' || constr[1] != 'A') {
2248                                 panic("lowering ASM 64bit output only supports '=A' currently");
2249                         }
2250                 }
2251         }
2252
2253         if (n_64bit_outs == 0)
2254                 return;
2255
2256         dbg_info          *dbgi       = get_irn_dbg_info(asmn);
2257         ir_node           *block      = get_nodes_block(asmn);
2258         ir_node           *mem        = get_ASM_mem(asmn);
2259         int                new_n_outs = 0;
2260         int                n_clobber  = get_ASM_n_clobbers(asmn);
2261         long              *proj_map   = ALLOCAN(long, n_outs);
2262         ident            **clobbers   = get_ASM_clobbers(asmn);
2263         ident             *asm_text   = get_ASM_text(asmn);
2264         ir_asm_constraint *new_outputs
2265                 = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2266         ir_node           *new_asm;
2267
2268         for (int o = 0; o < n_outs; ++o) {
2269                 const ir_asm_constraint *constraint = &output_constraints[o];
2270                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2271                         new_outputs[new_n_outs].pos        = constraint->pos;
2272                         new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2273                         new_outputs[new_n_outs].mode       = env->low_unsigned;
2274                         proj_map[o] = new_n_outs;
2275                         ++new_n_outs;
2276                         new_outputs[new_n_outs].pos        = constraint->pos;
2277                         new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2278                         if (constraint->mode == high_signed)
2279                                 new_outputs[new_n_outs].mode = env->low_signed;
2280                         else
2281                                 new_outputs[new_n_outs].mode = env->low_unsigned;
2282                         ++new_n_outs;
2283                 } else {
2284                         new_outputs[new_n_outs] = *constraint;
2285                         proj_map[o] = new_n_outs;
2286                         ++new_n_outs;
2287                 }
2288         }
2289         assert(new_n_outs == n_outs+(int)n_64bit_outs);
2290
2291         int       n_inputs = get_ASM_n_inputs(asmn);
2292         ir_node **new_ins  = ALLOCAN(ir_node*, n_inputs);
2293         for (int i = 0; i < n_inputs; ++i)
2294                 new_ins[i] = get_ASM_input(asmn, i);
2295
2296         new_asm = new_rd_ASM(dbgi, block, mem, n_inputs, new_ins, input_constraints,
2297                                                  new_n_outs, new_outputs, n_clobber, clobbers,
2298                                                  asm_text);
2299
2300         foreach_out_edge_safe(asmn, edge) {
2301                 ir_node *proj      = get_edge_src_irn(edge);
2302                 ir_mode *proj_mode = get_irn_mode(proj);
2303                 long     pn;
2304
2305                 if (!is_Proj(proj))
2306                         continue;
2307                 pn = get_Proj_proj(proj);
2308
2309                 if (pn < n_outs)
2310                         pn = proj_map[pn];
2311                 else
2312                         pn = new_n_outs + pn - n_outs;
2313
2314                 if (proj_mode == high_signed || proj_mode == high_unsigned) {
2315                         ir_mode *high_mode
2316                                 = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2317                         ir_node *np_low  = new_r_Proj(new_asm, env->low_unsigned, pn);
2318                         ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2319                         ir_set_dw_lowered(proj, np_low, np_high);
2320                 } else {
2321                         ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2322                         exchange(proj, np);
2323                 }
2324         }
2325 }
2326
2327 /**
2328  * Lower the builtin type to its higher part.
2329  *
2330  * @param mtp  the builtin type to lower
2331  *
2332  * @return the lowered type
2333  */
2334 static ir_type *lower_Builtin_type_high(ir_type *mtp)
2335 {
2336         ir_type *res;
2337         size_t   i;
2338         size_t   n_params;
2339         size_t   n_results;
2340         bool     must_be_lowered;
2341
2342         res = pmap_get(ir_type, lowered_builtin_type_high, mtp);
2343         if (res != NULL)
2344                 return res;
2345
2346         n_params        = get_method_n_params(mtp);
2347         n_results       = get_method_n_ress(mtp);
2348         must_be_lowered = false;
2349
2350         /* check for double word parameter */
2351         for (i = n_params; i > 0;) {
2352                 ir_type *tp = get_method_param_type(mtp, --i);
2353
2354                 if (is_Primitive_type(tp)) {
2355                         ir_mode *mode = get_type_mode(tp);
2356
2357                         if (mode == env->high_signed || mode == env->high_unsigned) {
2358                                 must_be_lowered = true;
2359                                 break;
2360                         }
2361                 }
2362         }
2363
2364         if (!must_be_lowered) {
2365                 set_type_link(mtp, NULL);
2366                 return mtp;
2367         }
2368
2369         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2370
2371         /* set param types and result types */
2372         for (i = 0; i < n_params; ++i) {
2373                 ir_type *tp = get_method_param_type(mtp, i);
2374
2375                 if (is_Primitive_type(tp)) {
2376                         ir_mode *mode = get_type_mode(tp);
2377
2378                         if (mode == env->high_signed) {
2379                                 if (env->params->little_endian) {
2380                                         set_method_param_type(res, i, tp_u);
2381                                 } else {
2382                                         set_method_param_type(res, i, tp_s);
2383                                 }
2384                         } else if (mode == env->high_unsigned) {
2385                                 set_method_param_type(res, i, tp_u);
2386                         } else {
2387                                 set_method_param_type(res, i, tp);
2388                         }
2389                 } else {
2390                         set_method_param_type(res, i, tp);
2391                 }
2392         }
2393         for (i = n_results = 0; i < n_results; ++i) {
2394                 ir_type *tp = get_method_res_type(mtp, i);
2395
2396                 if (is_Primitive_type(tp)) {
2397                         ir_mode *mode = get_type_mode(tp);
2398
2399                         if (mode == env->high_signed) {
2400                                 if (env->params->little_endian) {
2401                                         set_method_res_type(res, i, tp_u);
2402                                 } else {
2403                                         set_method_res_type(res, i, tp_s);
2404                                 }
2405                         } else if (mode == env->high_unsigned) {
2406                                 set_method_res_type(res, i, tp_u);
2407                         } else {
2408                                 set_method_res_type(res, i, tp);
2409                         }
2410                 } else {
2411                         set_method_res_type(res, i, tp);
2412                 }
2413         }
2414
2415         set_method_variadicity(res, get_method_variadicity(mtp));
2416         set_method_calling_convention(res, get_method_calling_convention(mtp));
2417         set_method_additional_properties(res, get_method_additional_properties(mtp));
2418
2419         pmap_insert(lowered_builtin_type_high, mtp, res);
2420         return res;
2421 }
2422
2423 /**
2424  * Lower the builtin type to its lower part.
2425  *
2426  * @param mtp  the builtin type to lower
2427  *
2428  * @return the lowered type
2429  */
2430 static ir_type *lower_Builtin_type_low(ir_type *mtp)
2431 {
2432         ir_type *res;
2433         size_t   i;
2434         size_t   n_params;
2435         size_t   n_results;
2436         bool     must_be_lowered;
2437
2438         res = pmap_get(ir_type, lowered_builtin_type_low, mtp);
2439         if (res != NULL)
2440                 return res;
2441
2442         n_params        = get_method_n_params(mtp);
2443         n_results       = get_method_n_ress(mtp);
2444         must_be_lowered = false;
2445
2446         /* check for double word parameter */
2447         for (i = n_params; i > 0;) {
2448                 ir_type *tp = get_method_param_type(mtp, --i);
2449
2450                 if (is_Primitive_type(tp)) {
2451                         ir_mode *mode = get_type_mode(tp);
2452
2453                         if (mode == env->high_signed || mode == env->high_unsigned) {
2454                                 must_be_lowered = true;
2455                                 break;
2456                         }
2457                 }
2458         }
2459
2460         if (!must_be_lowered) {
2461                 set_type_link(mtp, NULL);
2462                 return mtp;
2463         }
2464
2465         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2466
2467         /* set param types and result types */
2468         for (i = 0; i < n_params; ++i) {
2469                 ir_type *tp = get_method_param_type(mtp, i);
2470
2471                 if (is_Primitive_type(tp)) {
2472                         ir_mode *mode = get_type_mode(tp);
2473
2474                         if (mode == env->high_signed) {
2475                                 if (env->params->little_endian) {
2476                                         set_method_param_type(res, i, tp_s);
2477                                 } else {
2478                                         set_method_param_type(res, i, tp_u);
2479                                 }
2480                         } else if (mode == env->high_unsigned) {
2481                                 set_method_param_type(res, i, tp_u);
2482                         } else {
2483                                 set_method_param_type(res, i, tp);
2484                         }
2485                 } else {
2486                         set_method_param_type(res, i, tp);
2487                 }
2488         }
2489         for (i = 0; i < n_results; ++i) {
2490                 ir_type *tp = get_method_res_type(mtp, i);
2491
2492                 if (is_Primitive_type(tp)) {
2493                         ir_mode *mode = get_type_mode(tp);
2494
2495                         if (mode == env->high_signed) {
2496                                 if (env->params->little_endian) {
2497                                         set_method_res_type(res, i, tp_s);
2498                                 } else {
2499                                         set_method_res_type(res, i, tp_u);
2500                                 }
2501                         } else if (mode == env->high_unsigned) {
2502                                 set_method_res_type(res, i, tp_u);
2503                         } else {
2504                                 set_method_res_type(res, i, tp);
2505                         }
2506                 } else {
2507                         set_method_res_type(res, i, tp);
2508                 }
2509         }
2510
2511         set_method_variadicity(res, get_method_variadicity(mtp));
2512         set_method_calling_convention(res, get_method_calling_convention(mtp));
2513         set_method_additional_properties(res, get_method_additional_properties(mtp));
2514
2515         pmap_insert(lowered_builtin_type_low, mtp, res);
2516         return res;
2517 }
2518
2519 /**
2520  * lowers a builtin which reduces a 64bit value to a simple summary value
2521  * (popcount, ffs, ...)
2522  */
2523 static void lower_reduce_builtin(ir_node *builtin, ir_mode *mode)
2524 {
2525         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2526         ir_node         *operand      = get_Builtin_param(builtin, 0);
2527         ir_mode         *operand_mode = get_irn_mode(operand);
2528         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2529                 return;
2530
2531         {
2532         arch_allow_ifconv_func  allow_ifconv      = be_get_backend_param()->allow_ifconv;
2533         int                     arity             = get_irn_arity(builtin);
2534         dbg_info               *dbgi              = get_irn_dbg_info(builtin);
2535         ir_graph               *irg               = get_irn_irg(builtin);
2536         ir_type                *type              = get_Builtin_type(builtin);
2537         ir_type                *lowered_type_high = lower_Builtin_type_high(type);
2538         ir_type                *lowered_type_low  = lower_Builtin_type_low(type);
2539         ir_type                *result_type       = get_method_res_type(lowered_type_low, 0);
2540         ir_mode                *result_mode       = get_type_mode(result_type);
2541         ir_node                *block             = get_nodes_block(builtin);
2542         ir_node                *mem               = get_Builtin_mem(builtin);
2543         const lower64_entry_t  *entry             = get_node_entry(operand);
2544         ir_mode                *high_mode         = get_irn_mode(entry->high_word);
2545         ir_node                *in_high[1]        = {entry->high_word};
2546         ir_node                *in_low[1]         = {entry->low_word};
2547         ir_node                *res;
2548
2549         assert(is_NoMem(mem));
2550         assert(arity == 2);
2551
2552         switch (kind) {
2553         case ir_bk_ffs: {
2554                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2555                 ir_node               *zero_high      = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2556                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2557                 ir_node               *zero_result    = new_rd_Const(dbgi, irg, get_mode_null(result_mode));
2558                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2559                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero_high, ir_relation_equal);
2560                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2561                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2562                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2563                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2564                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2565                 ir_node               *mux_high       = new_rd_Mux(dbgi, block, cmp_high, high, zero_result, result_mode);
2566
2567                 if (! allow_ifconv(cmp_high, high, zero_result))
2568                         ir_nodeset_insert(&created_mux_nodes, mux_high);
2569
2570                 res = new_rd_Mux(dbgi, block, cmp_low, low, mux_high, result_mode);
2571
2572                 if (! allow_ifconv(cmp_low, low, mux_high))
2573                         ir_nodeset_insert(&created_mux_nodes, res);
2574                 break;
2575         }
2576         case ir_bk_clz: {
2577                 ir_node               *zero           = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2578                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero, ir_relation_equal);
2579                 ir_node               *clz_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2580                 ir_node               *high           = new_r_Proj(clz_high, result_mode, pn_Builtin_max+1);
2581                 ir_node               *clz_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2582                 ir_node               *low_proj       = new_r_Proj(clz_low, result_mode, pn_Builtin_max+1);
2583                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(mode));
2584                 ir_node               *low            = new_rd_Add(dbgi, block, low_proj, number_of_bits, result_mode);
2585
2586                 res = new_rd_Mux(dbgi, block, cmp_high, high, low, result_mode);
2587
2588                 if (! allow_ifconv(cmp_high, high, low))
2589                         ir_nodeset_insert(&created_mux_nodes, res);
2590                 break;
2591         }
2592         case ir_bk_ctz: {
2593                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2594                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2595                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2596                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2597                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2598                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2599                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2600                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2601
2602                 res = new_rd_Mux(dbgi, block, cmp_low, low, high, result_mode);
2603
2604                 if (! allow_ifconv(cmp_low, low, high))
2605                         ir_nodeset_insert(&created_mux_nodes, res);
2606                 break;
2607         }
2608         case ir_bk_popcount: {
2609                 ir_node               *popcount_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2610                 ir_node               *popcount_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2611                 ir_node               *high          = new_r_Proj(popcount_high, result_mode, pn_Builtin_max+1);
2612                 ir_node               *low           = new_r_Proj(popcount_low, result_mode, pn_Builtin_max+1);
2613
2614                 res = new_rd_Add(dbgi, block, high, low, result_mode);
2615                 break;
2616         }
2617         case ir_bk_parity: {
2618                 ir_node  *parity_high;
2619                 ir_node  *parity_low;
2620                 ir_node  *high;
2621                 ir_node  *low;
2622
2623                 assert(arity == 2);
2624
2625                 parity_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2626                 high        = new_r_Proj(parity_high, result_mode, pn_Builtin_max+1);
2627                 parity_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2628                 low         = new_r_Proj(parity_low, result_mode, pn_Builtin_max+1);
2629                 res         = new_rd_Eor(dbgi, block, high, low, result_mode);
2630                 break;
2631         }
2632         default:
2633                 panic("unexpected builtin");
2634         }
2635
2636         ir_node *const in[] = {
2637                 [pn_Builtin_M]       = mem,
2638                 [pn_Builtin_max + 1] = res,
2639         };
2640         turn_into_tuple(builtin, ARRAY_SIZE(in), in);
2641         }
2642 }
2643
2644 /**
2645  * lowers builtins performing arithmetic (bswap)
2646  */
2647 static void lower_arithmetic_builtin(ir_node *builtin, ir_mode *mode)
2648 {
2649         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2650         ir_node         *operand      = get_Builtin_param(builtin, 0);
2651         ir_mode         *operand_mode = get_irn_mode(operand);
2652         (void) mode;
2653         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2654                 return;
2655
2656         {
2657         dbg_info              *dbgi              = get_irn_dbg_info(builtin);
2658         ir_type               *type              = get_Builtin_type(builtin);
2659         ir_type               *lowered_type_high = lower_Builtin_type_high(type);
2660         ir_type               *lowered_type_low  = lower_Builtin_type_low(type);
2661         ir_node               *block             = get_nodes_block(builtin);
2662         ir_node               *mem               = get_Builtin_mem(builtin);
2663         const lower64_entry_t *entry             = get_node_entry(operand);
2664         ir_mode               *mode_high         = get_irn_mode(entry->high_word);
2665         ir_node               *res_high;
2666         ir_node               *res_low;
2667
2668         switch (kind) {
2669         case ir_bk_bswap: {
2670                 ir_node               *in_high[1] = { entry->high_word };
2671                 ir_node               *in_low[1]  = { entry->low_word };
2672                 ir_node               *swap_high  = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2673                 ir_node               *swap_low   = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2674                 ir_node               *high       = new_r_Proj(swap_high, mode_high, pn_Builtin_max+1);
2675                 ir_node               *low        = new_r_Proj(swap_low, env->low_unsigned, pn_Builtin_max+1);
2676                 if (mode_high == env->low_signed) {
2677                         res_high = new_rd_Conv(dbgi, block, low, env->low_signed);
2678                         res_low  = new_rd_Conv(dbgi, block, high, env->low_unsigned);
2679                 } else {
2680                         res_high = low;
2681                         res_low  = high;
2682                 }
2683                 break;
2684         }
2685         default:
2686                 panic("unexpected builtin");
2687         }
2688
2689         /* search result Proj */
2690         foreach_out_edge_safe(builtin, edge) {
2691                 ir_node *proj = get_edge_src_irn(edge);
2692                 if (!is_Proj(proj))
2693                         continue;
2694
2695                 if (get_Proj_proj(proj) == pn_Builtin_max+1) {
2696                         ir_set_dw_lowered(proj, res_low, res_high);
2697                 }
2698         }
2699         }
2700 }
2701
2702 /**
2703  * Lower double word builtins.
2704  */
2705 static void lower_Builtin(ir_node *builtin, ir_mode *mode)
2706 {
2707         ir_builtin_kind kind = get_Builtin_kind(builtin);
2708
2709         switch (kind) {
2710         case ir_bk_trap:
2711         case ir_bk_debugbreak:
2712         case ir_bk_return_address:
2713         case ir_bk_frame_address:
2714         case ir_bk_prefetch:
2715         case ir_bk_inport:
2716         case ir_bk_outport:
2717         case ir_bk_inner_trampoline:
2718                 /* Nothing to do. */
2719                 return;
2720         case ir_bk_bswap:
2721                 lower_arithmetic_builtin(builtin, mode);
2722                 return;
2723         case ir_bk_ffs:
2724         case ir_bk_clz:
2725         case ir_bk_ctz:
2726         case ir_bk_popcount:
2727         case ir_bk_parity:
2728                 lower_reduce_builtin(builtin, mode);
2729                 return;
2730         }
2731         panic("unknown builtin");
2732 }
2733
2734 /**
2735  * check for opcodes that must always be lowered.
2736  */
2737 static bool always_lower(unsigned code)
2738 {
2739         switch (code) {
2740         case iro_ASM:
2741         case iro_Builtin:
2742         case iro_Proj:
2743         case iro_Start:
2744         case iro_Call:
2745         case iro_Return:
2746         case iro_Cond:
2747         case iro_Switch:
2748         case iro_Conv:
2749         case iro_Sel:
2750                 return true;
2751         default:
2752                 return false;
2753         }
2754 }
2755
2756 /**
2757  * Compare two op_mode_entry_t's.
2758  */
2759 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2760 {
2761         const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2762         const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2763         (void) size;
2764
2765         return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2766 }
2767
2768 /**
2769  * Compare two conv_tp_entry_t's.
2770  */
2771 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2772 {
2773         const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2774         const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2775         (void) size;
2776
2777         return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2778 }
2779
2780 /**
2781  * Enter a lowering function into an ir_op.
2782  */
2783 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2784 {
2785         op->ops.generic = (op_func)func;
2786 }
2787
2788 /* Determine which modes need to be lowered */
2789 static void setup_modes(void)
2790 {
2791         unsigned           size_bits           = env->params->doubleword_size;
2792         ir_mode           *doubleword_signed   = NULL;
2793         ir_mode           *doubleword_unsigned = NULL;
2794         size_t             n_modes             = ir_get_n_modes();
2795         ir_mode_arithmetic arithmetic;
2796         unsigned           modulo_shift;
2797         size_t             i;
2798
2799         /* search for doubleword modes... */
2800         for (i = 0; i < n_modes; ++i) {
2801                 ir_mode *mode = ir_get_mode(i);
2802                 if (!mode_is_int(mode))
2803                         continue;
2804                 if (get_mode_size_bits(mode) != size_bits)
2805                         continue;
2806                 if (mode_is_signed(mode)) {
2807                         if (doubleword_signed != NULL) {
2808                                 /* sigh - the lowerer should really just lower all mode with
2809                                  * size_bits it finds. Unfortunately this required a bigger
2810                                  * rewrite. */
2811                                 panic("multiple double word signed modes found");
2812                         }
2813                         doubleword_signed = mode;
2814                 } else {
2815                         if (doubleword_unsigned != NULL) {
2816                                 /* sigh - the lowerer should really just lower all mode with
2817                                  * size_bits it finds. Unfortunately this required a bigger
2818                                  * rewrite. */
2819                                 panic("multiple double word unsigned modes found");
2820                         }
2821                         doubleword_unsigned = mode;
2822                 }
2823         }
2824         if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2825                 panic("Couldn't find doubleword modes");
2826         }
2827
2828         arithmetic   = get_mode_arithmetic(doubleword_signed);
2829         modulo_shift = get_mode_modulo_shift(doubleword_signed);
2830
2831         assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2832         assert(size_bits % 2 == 0);
2833         assert(get_mode_sign(doubleword_signed) == 1);
2834         assert(get_mode_sign(doubleword_unsigned) == 0);
2835         assert(get_mode_sort(doubleword_signed) == irms_int_number);
2836         assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2837         assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2838         assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2839
2840         /* try to guess a sensible modulo shift for the new mode.
2841          * (This is IMO another indication that this should really be a node
2842          *  attribute instead of a mode thing) */
2843         if (modulo_shift == size_bits) {
2844                 modulo_shift = modulo_shift / 2;
2845         } else if (modulo_shift == 0) {
2846                 /* fine */
2847         } else {
2848                 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2849         }
2850         size_bits /= 2;
2851
2852         /* produce lowered modes */
2853         env->high_signed   = doubleword_signed;
2854         env->high_unsigned = doubleword_unsigned;
2855         env->low_signed    = new_int_mode("WS", arithmetic, size_bits, 1,
2856                                           modulo_shift);
2857         env->low_unsigned  = new_int_mode("WU", arithmetic, size_bits, 0,
2858                                           modulo_shift);
2859 }
2860
2861 static void enqueue_preds(ir_node *node)
2862 {
2863         int arity = get_irn_arity(node);
2864         int i;
2865
2866         for (i = 0; i < arity; ++i) {
2867                 ir_node *pred = get_irn_n(node, i);
2868                 pdeq_putr(env->waitq, pred);
2869         }
2870 }
2871
2872 static void lower_node(ir_node *node)
2873 {
2874         int              arity;
2875         int              i;
2876         lower_dw_func    func;
2877         ir_op           *op;
2878         ir_mode         *mode;
2879         unsigned         idx;
2880         lower64_entry_t *entry;
2881
2882         if (irn_visited_else_mark(node))
2883                 return;
2884
2885         /* cycles are always broken at Phi and Block nodes. So we don't need special
2886          * magic in all the other lower functions */
2887         if (is_Block(node)) {
2888                 enqueue_preds(node);
2889                 return;
2890         } else if (is_Phi(node)) {
2891                 lower_Phi(node);
2892                 return;
2893         }
2894
2895         /* depth-first: descend into operands */
2896         if (!is_Block(node)) {
2897                 ir_node *block = get_nodes_block(node);
2898                 lower_node(block);
2899         }
2900
2901         if (!is_Cond(node)) {
2902                 arity = get_irn_arity(node);
2903                 for (i = 0; i < arity; ++i) {
2904                         ir_node *pred = get_irn_n(node, i);
2905                         lower_node(pred);
2906                 }
2907         }
2908
2909         op   = get_irn_op(node);
2910         func = (lower_dw_func) op->ops.generic;
2911         if (func == NULL)
2912                 return;
2913
2914         idx   = get_irn_idx(node);
2915         entry = idx < env->n_entries ? env->entries[idx] : NULL;
2916         if (entry != NULL || always_lower(get_irn_opcode(node))) {
2917                 mode = get_irn_op_mode(node);
2918                 if (mode == env->high_signed) {
2919                         mode = env->low_signed;
2920                 } else {
2921                         mode = env->low_unsigned;
2922                 }
2923                 DB((dbg, LEVEL_1, "  %+F\n", node));
2924                 func(node, mode);
2925         }
2926 }
2927
2928 static void clear_node_and_phi_links(ir_node *node, void *data)
2929 {
2930         (void) data;
2931         if (get_irn_mode(node) == mode_T) {
2932                 set_irn_link(node, node);
2933         } else {
2934                 set_irn_link(node, NULL);
2935         }
2936         if (is_Block(node))
2937                 set_Block_phis(node, NULL);
2938         else if (is_Phi(node))
2939                 set_Phi_next(node, NULL);
2940 }
2941
2942 static void lower_irg(ir_graph *irg)
2943 {
2944         ir_entity *ent;
2945         ir_type   *mtp;
2946         ir_type   *lowered_mtp;
2947         unsigned   n_idx;
2948
2949         obstack_init(&env->obst);
2950
2951         /* just here for debugging */
2952         current_ir_graph = irg;
2953         assure_edges(irg);
2954
2955         n_idx = get_irg_last_idx(irg);
2956         n_idx = n_idx + (n_idx >> 2);  /* add 25% */
2957         env->n_entries = n_idx;
2958         env->entries   = NEW_ARR_FZ(lower64_entry_t*, n_idx);
2959
2960         env->irg            = irg;
2961         env->flags          = 0;
2962
2963         ent = get_irg_entity(irg);
2964         mtp = get_entity_type(ent);
2965         lowered_mtp = lower_mtp(mtp);
2966
2967         if (lowered_mtp != mtp) {
2968                 set_entity_type(ent, lowered_mtp);
2969                 env->flags |= MUST_BE_LOWERED;
2970
2971                 fix_parameter_entities(irg, mtp);
2972         }
2973
2974         /* first step: link all nodes and allocate data */
2975         ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2976         visit_all_identities(irg, clear_node_and_phi_links, NULL);
2977         irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
2978
2979         if (env->flags & MUST_BE_LOWERED) {
2980                 size_t i;
2981                 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
2982                 inc_irg_visited(irg);
2983
2984                 assert(pdeq_empty(env->waitq));
2985                 pdeq_putr(env->waitq, get_irg_end(irg));
2986
2987                 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
2988                 while (!pdeq_empty(env->waitq)) {
2989                         ir_node *node = (ir_node*)pdeq_getl(env->waitq);
2990                         lower_node(node);
2991                 }
2992
2993                 /* we need to fixup phis */
2994                 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
2995                         ir_node *phi = env->lowered_phis[i];
2996                         fixup_phi(phi);
2997                 }
2998                 DEL_ARR_F(env->lowered_phis);
2999
3000
3001                 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
3002
3003                 if (env->flags & CF_CHANGED) {
3004                         /* control flow changed, dominance info is invalid */
3005                         clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
3006                 }
3007                 edges_deactivate(irg);
3008         }
3009
3010         ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3011
3012         DEL_ARR_F(env->entries);
3013         obstack_free(&env->obst, NULL);
3014 }
3015
3016 static const lwrdw_param_t *param;
3017
3018 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
3019 {
3020         assert(new_param != NULL);
3021         FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
3022
3023         param = new_param;
3024
3025         ir_clear_opcodes_generic_func();
3026         ir_register_dw_lower_function(op_ASM,     lower_ASM);
3027         ir_register_dw_lower_function(op_Add,     lower_binop);
3028         ir_register_dw_lower_function(op_And,     lower_And);
3029         ir_register_dw_lower_function(op_Bad,     lower_Bad);
3030         ir_register_dw_lower_function(op_Builtin, lower_Builtin);
3031         ir_register_dw_lower_function(op_Call,    lower_Call);
3032         ir_register_dw_lower_function(op_Cmp,     lower_Cmp);
3033         ir_register_dw_lower_function(op_Cond,    lower_Cond);
3034         ir_register_dw_lower_function(op_Const,   lower_Const);
3035         ir_register_dw_lower_function(op_Conv,    lower_Conv);
3036         ir_register_dw_lower_function(op_Div,     lower_Div);
3037         ir_register_dw_lower_function(op_Eor,     lower_Eor);
3038         ir_register_dw_lower_function(op_Load,    lower_Load);
3039         ir_register_dw_lower_function(op_Minus,   lower_unop);
3040         ir_register_dw_lower_function(op_Mod,     lower_Mod);
3041         ir_register_dw_lower_function(op_Mul,     lower_binop);
3042         ir_register_dw_lower_function(op_Mux,     lower_Mux);
3043         ir_register_dw_lower_function(op_Not,     lower_Not);
3044         ir_register_dw_lower_function(op_Or,      lower_Or);
3045         ir_register_dw_lower_function(op_Proj,    lower_Proj);
3046         ir_register_dw_lower_function(op_Return,  lower_Return);
3047         ir_register_dw_lower_function(op_Shl,     lower_Shl);
3048         ir_register_dw_lower_function(op_Shr,     lower_Shr);
3049         ir_register_dw_lower_function(op_Shrs,    lower_Shrs);
3050         ir_register_dw_lower_function(op_Start,   lower_Start);
3051         ir_register_dw_lower_function(op_Store,   lower_Store);
3052         ir_register_dw_lower_function(op_Sub,     lower_binop);
3053         ir_register_dw_lower_function(op_Switch,  lower_Switch);
3054         ir_register_dw_lower_function(op_Unknown, lower_Unknown);
3055 }
3056
3057 /**
3058  * Callback to lower only the Mux nodes we created.
3059  */
3060 static int lower_mux_cb(ir_node *mux)
3061 {
3062         return ir_nodeset_contains(&created_mux_nodes, mux);
3063 }
3064
3065 /*
3066  * Do the lowering.
3067  */
3068 void ir_lower_dw_ops(void)
3069 {
3070         lower_dw_env_t lenv;
3071         size_t      i, n;
3072
3073         memset(&lenv, 0, sizeof(lenv));
3074         lenv.params = param;
3075         env = &lenv;
3076
3077         setup_modes();
3078
3079         /* create the necessary maps */
3080         if (! intrinsic_fkt)
3081                 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
3082         if (! conv_types)
3083                 conv_types = new_set(cmp_conv_tp, 16);
3084         if (! lowered_type)
3085                 lowered_type = pmap_create();
3086         if (! lowered_builtin_type_low)
3087                 lowered_builtin_type_low = pmap_create();
3088         if (! lowered_builtin_type_high)
3089                 lowered_builtin_type_high = pmap_create();
3090
3091         /* create a primitive unsigned and signed type */
3092         if (! tp_u)
3093                 tp_u = get_type_for_mode(lenv.low_unsigned);
3094         if (! tp_s)
3095                 tp_s = get_type_for_mode(lenv.low_signed);
3096
3097         /* create method types for the created binop calls */
3098         if (! binop_tp_u) {
3099                 binop_tp_u = new_type_method(4, 2);
3100                 set_method_param_type(binop_tp_u, 0, tp_u);
3101                 set_method_param_type(binop_tp_u, 1, tp_u);
3102                 set_method_param_type(binop_tp_u, 2, tp_u);
3103                 set_method_param_type(binop_tp_u, 3, tp_u);
3104                 set_method_res_type(binop_tp_u, 0, tp_u);
3105                 set_method_res_type(binop_tp_u, 1, tp_u);
3106         }
3107         if (! binop_tp_s) {
3108                 binop_tp_s = new_type_method(4, 2);
3109                 if (env->params->little_endian) {
3110                         set_method_param_type(binop_tp_s, 0, tp_u);
3111                         set_method_param_type(binop_tp_s, 1, tp_s);
3112                         set_method_param_type(binop_tp_s, 2, tp_u);
3113                         set_method_param_type(binop_tp_s, 3, tp_s);
3114                         set_method_res_type(binop_tp_s, 0, tp_u);
3115                         set_method_res_type(binop_tp_s, 1, tp_s);
3116                 } else {
3117                         set_method_param_type(binop_tp_s, 0, tp_s);
3118                         set_method_param_type(binop_tp_s, 1, tp_u);
3119                         set_method_param_type(binop_tp_s, 2, tp_s);
3120                         set_method_param_type(binop_tp_s, 3, tp_u);
3121                         set_method_res_type(binop_tp_s, 0, tp_s);
3122                         set_method_res_type(binop_tp_s, 1, tp_u);
3123                 }
3124         }
3125         if (! unop_tp_u) {
3126                 unop_tp_u = new_type_method(2, 2);
3127                 set_method_param_type(unop_tp_u, 0, tp_u);
3128                 set_method_param_type(unop_tp_u, 1, tp_u);
3129                 set_method_res_type(unop_tp_u, 0, tp_u);
3130                 set_method_res_type(unop_tp_u, 1, tp_u);
3131         }
3132         if (! unop_tp_s) {
3133                 unop_tp_s = new_type_method(2, 2);
3134                 if (env->params->little_endian) {
3135                         set_method_param_type(unop_tp_s, 0, tp_u);
3136                         set_method_param_type(unop_tp_s, 1, tp_s);
3137                         set_method_res_type(unop_tp_s, 0, tp_u);
3138                         set_method_res_type(unop_tp_s, 1, tp_s);
3139                 } else {
3140                         set_method_param_type(unop_tp_s, 0, tp_s);
3141                         set_method_param_type(unop_tp_s, 1, tp_u);
3142                         set_method_res_type(unop_tp_s, 0, tp_s);
3143                         set_method_res_type(unop_tp_s, 1, tp_u);
3144                 }
3145         }
3146
3147         lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
3148         lenv.waitq         = new_pdeq();
3149         lenv.first_id      = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
3150         lenv.next_id       = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
3151
3152         irp_reserve_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3153         inc_master_type_visited();
3154         /* transform all graphs */
3155         for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
3156                 ir_graph *irg = get_irp_irg(i);
3157
3158                 ir_nodeset_init(&created_mux_nodes);
3159
3160                 lower_irg(irg);
3161
3162                 if (ir_nodeset_size(&created_mux_nodes) > 0)
3163                         lower_mux(irg, lower_mux_cb);
3164
3165                 ir_nodeset_destroy(&created_mux_nodes);
3166         }
3167         irp_free_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3168         del_pdeq(lenv.waitq);
3169
3170         env = NULL;
3171 }
3172
3173 /* Default implementation. */
3174 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
3175                                     const ir_mode *imode, const ir_mode *omode,
3176                                     void *context)
3177 {
3178         char buf[64];
3179         ident *id;
3180         ir_entity *ent;
3181         (void) context;
3182
3183         if (imode == omode) {
3184                 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
3185         } else {
3186                 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
3187                         get_mode_name(imode), get_mode_name(omode));
3188         }
3189         id = new_id_from_str(buf);
3190
3191         ent = new_entity(get_glob_type(), id, method);
3192         set_entity_ld_ident(ent, get_entity_ident(ent));
3193         set_entity_visibility(ent, ir_visibility_external);
3194         return ent;
3195 }